As a Chinese user, I can say that many people use Kimi, even though I personally don’t use it much. China’s open-source strategy has many significant effects—not only because it aligns with the spirit of open source. For domestic Chinese companies, it also prevents startups from making reckless investments to develop mediocre models. Instead, everyone is pushed to start from a relatively high baseline. Of course, many small companies in the U.S., Japan, and Europe are also building on Qwen. Kimi is similar: before DeepSeek and others emerged, their model quality was pretty bad. Once the open-source strategy was set, these companies had no choice but to adjust their product lines and development approaches to improve their models.
Moreover, the ultimate competition between models will eventually become a competition over energy. China’s open-source models have major advantages in energy consumption, and China itself has a huge advantage in energy resources. They may not necessarily outperform the U.S., but they probably won’t fall too far behind either.
There’s a lot of indications that we’re currently brute forcing these models. There’s honestly not a reason they have to be 1T parameters and cost an insane amount to train and run on inference.
What we’re going to see is as energy becomes a problem; they’ll simply shift to more effective and efficient architectures on both physical hardware and model design. I suspect they can also simply charge more for the service, which reduces usage for senseless applications.
There are also elements of stock price hype and geopolitical competition involved.
The major U.S. tech giants are all tied to the same bandwagon — they have to maintain this cycle:
buy chips → build data centers → release new models → buy more chips.
It might only stop once the electricity problem becomes truly unsustainable.
Of course, I don’t fully understand the specific situation in the U.S.,
but I even feel that one day they might flee the U.S. altogether and move to the Middle East to secure resources.
I suspect that the OpenRouter result originates from a quantized hosting provider. The difference compared to the direct API call from Moonshot is striking, almost like night and day. It creates a peculiar user and developer experience since OpenRouter enforces quantization restrictions only at the API level, rather than at the account settings level.
Love seeing this benchmark become more iconic with each new model release. Still in disbelief at the GPT-5 variants' performance in comparison but its cool to see the new open source models get more ambitious with their attempts.
Dataset contamination alone won't get them good-looking SVG pelicans on bicycles though, they'll have to either cheat this particular question specifically or train it to make vector illustrations in general. At which point it can be easily swapped for another problem that wasn't in the data.
they can have some cheap workers make about 10 pelicans by hand in svg, fuzz them to generate thousands of variations and throw it in their training pool. don't need to 'get good at svgs' by any means.
I actually prefer ascii art diagrams as a benchmark for visual thinking, since it requires 2 stages,
Like svg, and also can test imaginative repurposing of text elements.
It started as a joke, but over time performance on this one weirdly appears to correlate to how good the models are generally. I'm not entirely sure why!
I'm not saying its objective or quantitative, but I do think its an interesting task because it would be challenging for most humans to come up with a good design of a pelican riding a bicycle.
There are many reports of CLI AI tools displaying words that humans express when they are frustrated and about to give up. Just what they have been trained on. That does not mean they have emotions. And "deleting the whole codebase" sounds more interesting, but I assume is the same thing. "Frustrated" words lead to frustrated actions. Does not mean the LLM was frustrated. Just that in its training data those things happened so it copied them in that situation.
The difference is, people and animals have a body, nerve system and in general those mushy things we think are responsible for emotions.
Computers don't have any of that. And LLM's in particular neither. They were trained to simulate human text responses, that's all. How to get from there to emotions - where is the connection?
Don't confuse the medium with the picture it represents.
Porn is pornographic, whether it is a photo or an oil painting.
Feelings are feelings, whether they're felt by a squishy meat brain or a perfect atom-by-atom simulation of one in a computer. Or a less-than-perfect simulation of one. Or just a vaguely similar system that is largely indistinguishable from it, as observed from the outside.
Individual nerve cells don't have emotions! Ten wired together don't either. Or one hundred, or a thousand... by extension you don't have any feelings either.
I think its cool and useful precisely because its not trying to correlate intelligence. It's a weird kind of niche thing that at least intuitively feels useful for judging llms in particular.
I'd much prefer a test which measures my cholesterol than one that would tell me whether I am an elf or not!
I've been under the impression most inference engines aren't fully deterministic with a temperature of 0 as some of the initial seed values can vary.
Note: I haven't tested this nor have I played with seed values. IIRC the inference engines I used support an explicit seed value, that is randomized by default.
If you want to do it at home, ik_llama.cpp has some performance optimizations that make it semi-practical to run a model of this size on a server with lots of memory bandwidth and a GPU or two for offload. You can get 6-10 tok/s with modest hardware workstation hardware. Thinking chews up a lot of tokens though, so it will be a slog.
Hi Simon. I have a Xeon W5-3435X with a 768GB of DDR5 across 8 channels, iirc it's running at 5800MT/s. It also has 7x A4000s, water cooled to pack them into a desktop case. Very much a compromise build, and I wouldn't recommend Xeon sapphire rapids because the memory bandwidth you get in practice is less than half of what you'd calculate from the specs. If I did it again, I'd build an EPYC machine with 12 channels of DDR5 and put in a single rtx 6000 pro blackwell. That'd be a lot easier and probably a lot faster.
There's a really good thread on level1techs about running DeepSeek at home, and everything there more-or-less applies to Kimi K2.
It's good to see more competition, and open source, but I'd be much more excited to see what level of coding and reasoning performance can be wrung out of a much smaller LLM + agent as opposed to a trillion parameter one. The ideal case would be something that can be run locally, or at least on a modest/inexpensive cluster.
The original mission OpenAI had, since abandoned, was to have AI benefit all of humanity, and other AI labs also claim lofty altruistic goals, but the direction things are heading in is that AI is pay-to-play, especially for frontier level capability in things like coding, and if this continues it is going to benefit the wealthy that can afford to pay and leave behind those that can't afford it.
> I'd be much more excited to see what level of coding and reasoning performance can be wrung out of a much smaller LLM + agent
Well, I think you are seeing that already? It's not like these models don't exist and they did not try to make them good, it's just that the results are not super great.
And why would they be? Why would the good models (that are barely okay at coding) be big, if it was currently possible to build good models, that are small?
Of course, new ideas will be found and this dynamic may drastically change in the future, but there is no reason to assume that people who work on small models find great optimizations that frontier models makers, who are very interested in efficient models, have not considered already.
Sure, but that's the point ... today's locally runnable models are a long way behind SOTA capability, so it'd be nice to see more research and experimentation in that direction. Maybe a zoo of highly specialized small models + agents for S/W development - one for planning, one for coding, etc?
If I understand transformers properly, this is unlikely to work. The whole point of “Large” Language Models is that you primarily make them better by making them larger, and when you do so, they get better at both general and specific tasks (so there isn’t a way to sacrifice generality but keep specific skills when training a small models).
I know a lot of people want this (Apple really really wants this and is pouring money into it) but just because we want something doesn’t mean it will happen, especially if it goes against the main idea behind the current AI wave.
I’d love to be wrong about this, but I’m pretty sure this is at least mostly right.
I think this is a description of how things are today, but not an inherent property of how the models are built. Over the last year or so the trend seems to be moving from “more data” to “better data”. And I think in most narrow domains (which, to be clear, general coding agent is not!) it’s possible to train a smaller, specialized model reaching the performance of a much larger generic model.
Actually there are ways you might get on device models to perform well. It is all about finding ways to have a smaller number of weights work efficiently.
One way is reusing weights in multiple decoders layers. This works and is used in many on-device models.
It is likely that we can get pretty high performance with this method. You can also combine this with low parameter ways to create overlapped behavior on the same weights as well, people had done LORA on top of shared weights.
Personally I think there are a lot of potential ways that you can cause the same weights to exhibit "overloaded" behaviour in multiple places in the same decoder stack.
Edit: I believe this method is used a bit for models targeted for the phone. I don't think we have seen significant work on people targeting say a 3090/4090 or similar inference compute size.
The issue isn't even 'quality' per se (for many tasks a small model would do fine), its for "agentic" workflows it _quickly_ runs out of context. Even 32GB VRAM is really very limiting.
And when I mean agentic, i mean something even like this - 'book a table from my emails', which involves looking at 5k+ tokens of emails, 5k tokens of search results, then confirming with the user etc. It's just not feasible on most hardware right now - even if the models are 1-2GB, you'll burn thru the rest in context so quickly.
Yeah - the whole business model of companies like OpenAI and Anthropic, at least at the moment, seems to be that the models are so big that you need to run them in the cloud with metered access. Maybe that could change in the future to sale or annual licence business model if running locally became possible.
I think scale helps for general tasks where the breadth of capability may be needed, but it's not so clear that this needed for narrow verticals, especially something like coding (knowing how to fix car engines, or distinguish 100 breeds of dog is not of much use!).
> the whole business model of companies like OpenAI and Anthropic, at least at the moment, seems to be that the models are so big that you need to run them in the cloud with metered access.
That's not a business model choice, though. That's a reality of running SOTA models.
If OpenAI or Anthropic could squeeze the same output out of smaller GPUs and servers they'd be doing it for themselves. It would cut their datacenter spend dramatically.
> If OpenAI or Anthropic could squeeze the same output out of smaller GPUs and servers they'd be doing it for themselves.
First, they do this; that's why they release models at different price points. It's also why GPT-5 tries auto-routing requests to the most cost-effective model.
Second, be careful about considering the incentives of these companies. They all act as if they're in an existential race to deliver 'the' best model; the winner-take-all model justifies their collective trillion dollar-ish valuation. In that race, delivering 97% of the performance at 10% of the cost is a distraction.
No I don’t think it’s a business model thing, I’m saying it may be a technical limitation of LLMs themselves. Like, that that there’s no way to “order a la carte” from the training process, you either get the buffet or nothing, no matter how hungry you feel.
> today's locally runnable models are a long way behind SOTA capability
SOTA models are larger than what can be run locally, though.
Obviously we'd all like to see smaller models perform better, but there's no reason to believe that there's a hidden secret to making small, locally-runnable models perform at the same level as Claude and OpenAI SOTA models. If there was, Anthropic and OpenAI would be doing it.
There's research happening and progress being made at every model size.
The point is still valid. If the big companies could save money running multiple small specialised models on cheap hardware, they wouldn't be spending billions on the highest spec GPUs.
You want more research on small language models? You're confused. There is already WAY more research done on small language models (SLM) than big ones. Why? Because it's easy. It only takes a moderate workstation to train an SLM. So every curious Masters student and motivated undergrad is doing this. Lots of PhD research is done on SLM because the hardware to train big models is stupidly expensive, even for many well-funded research labs. If you read Arxiv papers (not just the flashy ones published by companies with PR budgets) most of the research is done on 7B parameter models. Heck, some NeurIPS papers (extremely competitive prestigious) from _this year_ are being done on 1.5B parameter models.
Lack of research is not the problem. It's fundamental limitations of the technology. I'm not gonna say "there's only so much smarts you can cram into a 7B parameter model" - because we don't know that yet for sure. But we do know, without a sliver of a doubt, that it's VASTLY EASIER to cram a smarts into a 70B parameter model than a 7B param model.
> In LLMs, we will have bigger weights vs test-time compute tradeoffs. A smaller model can get "there" but it will take longer.
Assuming both are SOTA, a smaller model can't produce the same results as a larger model by giving it infinite time. Larger models inherently have more room for training more information into the model.
No amount of test-retry cycle can overcome all of those limits. The smaller models will just go in circles.
I even get the larger hosted models stuck chasing their own tail and going in circles all the time.
It's true that to train more information into the model you need more trainable parameters, but when people ask for small models, they usually mean models that run at acceptable speeds on their hardware. Techniques like mixture-of-experts allow increasing the number of trainable parameters without requiring more FLOPs, so they're large in one sense but small in another.
And you don't necessarily need to train all information into the model, you can also use tool calls to inject it into the context. A small model that can make lots of tool calls and process the resulting large context could obtain the same answer that a larger model would pull directly out of its weights.
Almost all training data are on the internet. As long as the small model has enough agentic browsing ability, given it enough time it will retrieve the data from the internet.
I have spent the last 2.5 years living like a monk to maintain an app across all paid LLM providers and llama.cpp.
I wish this was true.
It isn't.
"In algorithms, we have space vs time tradeoffs, therefore a small LLM can get there with more time" is the same sort of "not even wrong" we all smile about us HNers doing when we try applying SWE-thought to subjects that aren't CS.
What you're suggesting amounts to "monkeys on typewriters will write entire works of Shakespeare eventually" - neither in practice, nor in theory, is this a technical claim, or something observable, or even stood up as a one-off misleading demo once.
If "not even wrong" is more wrong than wrong, then is 'not even right" more right than right.
To answer you directly, a smaller SOTA reasoning model with a table of facts can rederive relationships given more time than a bigger model which encoded those relationships implicitly.
This doesn't work like that. An analogy would be giving a 5 year old a task that requires the understanding of the world of an 18 year old. It doesn't matter whether you give that child 5 minutes or 10 hours, they won't be capable of solving it.
I think the question of what can be achieved with a small model comes down to what needs knowledge vs what needs experience. A small model can use tools like RAG if it is just missing knowledge, but it seems hard to avoid training/parameters where experience is needed - knowing how to perceive then act.
There is obviously also some amount (maybe a lot) of core knowledge and capability needed even to be able to ask the right questions and utilize the answers.
It is not clear that a simple/small model with inference running on home hardware is energy or cost efficient compared to the scaled up inference of a large model with batch processing. There are dozens of optimizations possible when splitting an LLM on multiple tiny components on separate accelerator units and when one handles kv cache optimization at the data center level; these are simply not possible at home and would be a waste of effort and energy until you serve thousands to millions of requests in parallel.
"open source" means there should be a script that downloads all the training materials and then spins up a pipeline that trains end to end.
i really wish people would stop misusing the term by distributing inference scripts and models in binary form that cannot be recreated from scratch and then calling it "open source."
They'd have to publish or link the training data, which is full of copyrighted material. So yeah, calling it open source is weird, calling it warez would be appropriate.
it still doesn't sit right. sure it's different in terms of mutability from say, compiled software programs, but it still remains not end to end reproducible and available for inspection.
these words had meaning long before "model land" became a thing. overloading them is just confusing for everyone.
It's not confusing, no one is really confused except the people upset that the meaning is different in a different context.
On top of that, in many cases a company/group/whoever can't even reproduce the model themselves. There are lots of sources of non-determinism even if folks are doing things in a very buttoned up manner. And, when you are training on trillions of tokens, you are likely training on some awful sounding stuff - "Facebook is trained llama 4 on nazi propaganda!" is not what they want to see published.
i disagree. words matter. the whole point of open source is that anyone can look and see exactly how the sausage is made. that is the point. that is why the word "open" is used.
...and sure, compiling gcc is nondeterministic too, but i can still inspect the complete source from where it comes because it is open source, which means that all of the source materials are available for inspection.
The point of open source in software is as you say. It's just not the same thing though. Using words and phrases differently in different fields is common.
Yeah, but "open weights" never seems to have taken off as a better description, and even if you did have the training data + recipe, the compute cost makes training it yourself totally impractical.
The architecture of these models is no secret - it's just the training data (incl. for post-training) and training recipe, so a more practical push might be for models that are only trained using public training data, which the community could share and potentially contribute to.
I'd agree but we're beyond hopelessly idealistic. That sort of approach only helps your competition who will use it to build a closed product and doesn't give anything of worth to people who want to actually use the model because they have no means to train it. Hell most people can barely scrape up enough hardware to even run inference.
Reproducing models is also not very ecological in when it comes down to it, do we really all need to redo the training that takes absurd amounts of power just to prove that it works? At least change the dataset to try and get a better result and provide another datapoint, but most people don't have the knowhow for it anyway.
Nvidia does try this approach sometimes funnily enough, they provide cool results with no model in hopes of getting people to buy their rented compute and their latest training platform as a service...
Yes, I am also super interested in cutting the size of models.
However, in a few years today’s large models will run locally anyhow.
My home computer had 16KB RAM in 1983. My $20K research workstation had 192MB of RAM in 1995. Now my $2K laptop has 32GB.
There is still such incredible pressure on hardware development that you can be confident that today’s SOTA models will be running at home before too long, even without ML architecture breakthroughs. Hopefully we will get both.
Edit: the 90’s were exciting for compute per dollar improvements. That expensive Sun SPARC workstation I started my PhD with was obsolete three years later, crushed by a much faster $1K Intel Linux beige box. Linux installed from floppies…
> My home computer had 16KB RAM in 1983. My $20K research workstation had 192MB of RAM in 1995. Now my $2K laptop has 32GB.
You’ve picked the wrong end of the curve there. Moore’s law was alive and kicking in the 90s. Every 1-3 years brought an order of magnitude better CPU and memory. Then we hit a wall. Measuring from the 2000s is more accurate.
My desktop had 4GB of RAM in 2005. In 20 years it’s gone up by a factor of 8, but only by a factor of 2 in the past 10 years.
I can kind of uncomfortably run a 24B parameter model on my MacBook Pro. That’s something like 50-200X smaller (depending on quantization) than a 1T parameter model.
We’re a _long_ way from having enough RAM (let alone RAM in the GPU) for this size of model. If the 8x / 20 years holds, we’re talking 40-60 years. If 2X / 10 years holds, we’re talking considerably longer. If the curve continues to flatten, it’s even longer.
Not to dampen anyone’s enthusiasm, but let’s be realistic about hardware improvements in the 2010s and 2020s. Smaller models will remain interesting for a very long time.
Moore’s Law is about transistor density, not RAM in workstations. But yes, density is not doubling every two years any more.
RAM growth slowed in laptops and workstations because we hit diminishing returns for normal-people applications. If local LLM applications are in demand, RAM will grow again.
What i do not understand is why we are not seeing specialized models that go down to single experts.
I do not need models that know how to program in Python, Rust, ... when i only use Go and Html. So we are we not seeing models that have very specialized experts, where for instance:
* General interpreter model, that holds context/memory
* Go Model
* Html model if there is space in memory.
* SQL model if there is space in memory.
If there is no space, the GIM swamp out the Go model, for the HTML model, depending on where it is in Agent tasks or Edit/Ask code its overviewing.
Because the models are going to be very small, switching in and out of memory will be ultra fast But most of the time we get very big Expert models, that still are very generalized over a entire field.
This can then be extended that if you have the memory, models combine their output with tasks... Maybe i am just too much of a noob in the field of understanding how LLMs work, but it feels like people are too often running after large models that companies like Anthropic/OpenAI etc deploy. I understand why those big companies use insane big models. They have the money to load them up over a cluster, have the fast interconnect, and for them its more efficient.
But from the bits and pieces that i see, people are more and more going to tons of small 1 a 2B models to produce better results. See my argument above. Like i said, never really gone beyond paying for my CoPilot subscription and running a bit of Ollama at home (don't have the time for the big stuff).
I think one of the issues is that LLMs can't have a "Go" model and an "HTML model". I mean, they can but what would that contain? It's not the language-specific features that make models large.
When models work on your code base, they do not "see" things like this, which is why they can go through an entire code base with variable names they have never seen before, function signatures they have never seen before, and directory structures that have never seen before and not have a problem.
You need that "this is a variable, which is being passed to a function which recursively does ..." part. This is not something language specific, it's the high level understanding of how languages and systems operate. A variable is a variable whether in JavaScript or C++ and LLMs can "see" it as such. The details are different but it's that layer of "this is a software interface", "this is a function pointer" is outside of the "Go" or "Python" or "C#" model.
I don't know how large the main model would have to be vs. the specialized models in order to pick this dynamic up.
You wont win much performance with a specific coding language tokenizer/vocabulary, everything else benefits from a larger model size. You can get distilled models that will out-perform or compete with your single domain coding model
With these things it’s always both at the same time: these super grandiose SOTA models are only making improvements mostly because of optimizations, and they’re just scaling our as far as they can.
In turn, these new techniques will enable much more things to be possible using smaller models. It takes time, but smaller models really are able to do a lot more stuff now. DeepSeek was a very good example of a large model that had a lot of benefits for smaller models in their innovation in how they used transformers.
Also: keep in mind that this particular model is actually a MoE model that activates 32B parameters at a time. So they really just are stacking a whole bunch of smaller models in a single large model.
I used to be obsessed with what's the smartest LLM, until I tried actually using them for some tasks and realized that the smaller models did the same task way faster.
So I switched my focus from "what's the smartest model" to "what's the smallest one that can do my task?"
With that lens, "scores high on general intelligence benchmarks" actually becomes a measure of how overqualified the model is, and how much time, money and energy you are wasting.
> The ideal case would be something that can be run locally, or at least on a modest/inexpensive cluster.
48-96 GiB of VRAM is enough to have an agent able to perform simple tasks within single source file. That's the sad truth. If you need more your only options are the cloud or somehow getting access to 512+ GiB
I think there is a lot of progress on efficient useful models recently.
I've seen GLM-4.6 getting mention for good coding results from a model that's much smaller than Kimi (~350b params) and seen it speculated that Windsurf based their new model on it.
This Kimi release is natively INT4, with quantization-aware training. If that works--if you can get really good results from four-bit parameters--it seems like a really useful tool for any model creator wanting efficient inference.
DeepSeek's v3.2-Exp uses their sparse attention technique to make longer-context training and inference more efficient. Its output's being priced at 60% less than v3.1 (though that's an imperfect indicator of efficiency). They've also quietly made 'thinking' mode need fewer tokens since R1, helping cost and latency.
And though it's on the proprietary side, Haiku 4.5 approaching Sonnet 4 coding capability (at least on benches Anthropic released) also suggests legitimately useful models can be much smaller than the big ones.
There's not yet a model at the level of any of the above that's practical for many people to run locally, though I think "efficient to run + open so competing inference providers can run it" is real progress.
More important it seems like there's a good trendline towards efficiency, and a bunch of techniques are being researched and tested that, when used together, could make for efficient higher-quality models.
>The ideal case would be something that can be run locally, or at least on a modest/inexpensive cluster.
It's obviously valuable, so it should be coming. I expect 2 trends:
- Local GPU/NPU will have a for-LLM version that has 50-100GB VRAM and runs MXFP4 etc.
- Distillation will come for reasoning coding agents, probably one for each tech stack (LAMP, Android app, AWS, etc.)x business domain (gaming, social, finance, etc.)
Even if pay-to-play companies like moonshootai help to pay less.
You can run previous kimi k2 non-thinking model e.g. on groq with 720tok/s and for $1/$3 for million input/output tokens. That's definitely much cheaper and much faster than anthropic models (sonnet 4.5: 60tok/s, $3/$15)
I think it’s going to be a while before we see small models (defined roughly as “runnable on reasonable consumer hardware”) do a good job at general coding tasks. It’s a very broad area! You can do some specific tasks reasonably well (eg I distilled a toy git helper you can run locally here https://github.com/distil-labs/gitara), but “coding” is such a big thing that you really need a lot of knowledge to do it well.
I think that's where prompt engineering would be needed. Bigger models produce good output even with ambiguous prompts. Getting similar output from smaller models is art,
Software development is one of the areas where LLMs really are useful, whether that's vibe coding disposable software, or more structured use for serious development.
I've been a developer for 40+ years, and very good at it, but for some tasks it's not about experience or overcoming complexity - just a bunch of grunt work that needs to come together. The other day I vibe coded a prototype app, just for one-time demo use, in less than 15 min that probably would have taken a week to write by hand, assuming one was already familiar with the tech stack.
Developing is fun, and a brain is a terrible thing to waste, but today not using LLMs where appropriate for coding doesn't make any sense if you value your time whatsoever.
That's going to depend on how small the model can be made, and how much you are using it.
If we assume that running locally meant running on a 500W consumer GPU, then the electricity cost to run this non-stop 8 hours a day for 20 days a month (i.e. "business hours") would be around $10-20.
This is about the same as OpenAI or Anthropics $20/mo plans, but for all day coding you would want their $100 or $200/mo plans, and even these will throttle you and/or require you to switch to metered pricing when you hit plan limits.
Four independent Chinese companies released extremely good open source models in the past few months (DeepSeek, Qwen/Alibaba, Kimi/Moonshot, GLM/Z.ai). No American or European companies are doing that, including titans like Meta. What gives?
I like Qwen 235 quite a bit too, and I generally agree with your sentiment, but this was a very large American open source model.
Unless we're getting into the complications on what "open source" model actually means, in which case I have no clue if these are just open weight or what.
The Chinese are doing it because they don't have access to enough of the latest GPUs to run their own models. Americans aren't doing this because they need to recoup the cost of their massive GPU investments.
Why is inference less attainable when it technically requires less GPU processing to run? Kimi has a chat app on their page using K2 so they must have figured out inference to some extent.
Inference is usually less gpu-compute heavy, but much more gpu-vram heavy pound-for-pound compared to training. General rule of thumb is that you need 20x more vram for training a model with X params, than for inference for that same size model. So assuming batch size b, then serving more than 20*b users would tilt vram use on the side of inference.
This isn't really accurate; it's an extremely rough rule of thumb and ignores a lot of stuff. But it's important to point out that inference is quickly adding to costs for all AI companies. Deepseek claims that they used $5.6mil to train Deepseek R1; that's about 10-20 trillion tokens at their current pricing- or 1 million users sending just 100 requests at full context size.
At ECAI conference last week there was a panel discussion and someone had a great quote, "in Europe we are in the golden age of AI regulation, while the US and China are in the actual golden age of AI".
"Who could've predicted?" as a sarcastic response to someone's stupid actions leading to entirely predictable consequences is probably as old as sarcasm itself.
Honestly, do we need to? If the Chinese release SOTA open source models, why should we invest a ton just to have another one? We can just use theirs, that's the beauty of open source.
For the vast majority, they're not "open source" they're "open weights". They don't release the training data or training code / configs.
It's kind of like releasing a 3d scene rendered to a JPG vs actually providing someone with the assets.
You can still use it, and it's possible to fine-tune it, but it's not really the same. There's tremendous soft power in deciding LLM alignment and material emphasis. As these things become more incorporated into education, for instance, the ability to frame "we don't talk about ba sing se" issues are going to be tremendously powerful.
Europe is in perpetual shambles so I wouldn’t even ask them for input on anything, really. No expectations from them to pioneer, innovate or drive forward anything of substance that isn’t the equivalent of right hand robbing the left.
* Our satellites are giving us by far the best understanding of our universe, capturing one third of the visible sky in incredible detail - just check out this mission update video if you want your mind blown: https://www.youtube.com/watch?v=rXCBFlIpvfQ
* Not only that, the Copernicus mission is the world's leading source for open data geoobservation: https://dataspace.copernicus.eu/
* We've given the world mRNA vaccines to solve the Covid crisis and GLP-1 antagonists to solve the obesity crisis.
* CERN and is figuring out questions about the fundamental nature of the universe, with the LHC being by far the largest particle accelerator in the world, an engineering precision feat that couldn't have been accomplished anywhere else.
Pioneering, innovation and drive forward isn't just about the latest tech fad. It's about fundamental research on how our universe works. Everyone else is downstream of us.
The answer is simply that no one would pay to use them for a number of reasons including privacy. They have to give them away and put up some semblance of openness. No option really.
I know first hand companies paying them. Chinese internal software market is gigantic. Full of companies and startups that have barely made into a single publication in the west.
Of course they are paying them. That’s not my point. My point is this is the only way for them to gain market share and they need Western users to train future models. They have to give them away. I’d be shocked if compute costs are not heavily subsidized by CCP.
But the CCP only has access to the US market because they joined the WTO, but when they joined the WTO they signed a treaty that they wouldn't do things like that.
I don’t think there’s any privacy that OpenAI or Anthropic are giving you that DeepSeek isn’t giving you. ChatGPT usage logs were held by court order at one point.
It’s true that DeepSeek won’t give you reliable info on Tiananmen Square but I would argue that’s a very rare use case in practice. Most people will be writing boilerplate code or summarizing mundane emails.
Also, the Meta AI 'team' is currently retooling so they can put something together with a handful of Zuck-picked experts making $100m+ each rather than hundreds making ~$1m each.
Love their nonsense excuse they they are trying to protect us from misuse of "superintelligence".
>“We believe the benefits of superintelligence should be shared with the world as broadly as possible. That said, superintelligence will raise novel safety concerns. We’ll need to be rigorous about mitigating these risks and careful about what we choose to open source.” -Mark Zuckerberg
Meta has shown us daily that they have no interest in protecting anything but their profits. They certainly don't intend to protect people from the harm their technology may do.
They just know that saying "this is profitable enough for us to keep it proprietary and restrict it to our own paid ecosystem" will make the enthusiasts running local Llama models mad at them.
Kimi K2 Thinking, MiniMax M2 Interleaved Thinking: open models are reaching, or have reached, frontier territory. We now have GPT and Claude Sonnet capable at home, as they are open-weight. Around this time last year, we had the DeepSeek moment, Now is the time for another moment.
Is the price here correct? https://openrouter.ai/moonshotai/kimi-k2-thinking
Would be $0,60 for input and $2,50 for 1 million output tokens. If the model is really that good it's 4x cheaper than comparable models. It's hosted at a loss or the others have a huge margin? I might miss something here.
Would love some expert opinion :)
Somehow that article totally ignored the insane pricing of cached input tokens set by Anthropic and OpenAI. For agentic coding, typically 90~95% of the inference cost is attributed to cached input tokens, and a scrappy China company can do it almost for free: https://api-docs.deepseek.com/news/news0802
Yes, you may consider that opensource models hosted over Openrouter are charging about bare hardware costs, where in practice some providers there may run on subsidized hardware even, so there is money to be made.
I am sure they cherry-picked the examples but still, wow. Having spent a considerable amount of time trying to introduce OSS models in my workflows I am fully aware of their short comings. Even frontier models would struggle with such outputs (unless you lead the way, help break down things and maybe even use sub-agents).
Very impressed with the progress. Keeps me excited about what’s to come next!
I like Kimi too, but they definitely have some benchmark contamination: the blog post shows a substantial comparative drop in swebench verified vs open tests. I throw no shade - releasing these open weights is a service to humanity; really amazing.
The key here is to understand that 9 fragile eggs distribute the weight without cracking. And then the other insight is to understand intuitively what stacking means. Where arranging things around certain objects doesn't make any sense.
Maybe a dumb question but: what is a "reasoning model"?
I think I get that "reasoning" in this context refers to dynamically budgeting scratchpad tokens that aren't intended as the main response body. But can't any model do that, and it's just part of the system prompt, or more generally, the conversation scaffold that is being written to.
Or does a "reasoning model" specifically refer to models whose "post training" / "fine tuning" / "rlhf" laps have been run against those sorts of prompts rather than simpler user-assistant-user-assistant back and forths?
EG, a base model becomes "a reasoning model" after so much experience in the reasoning mines.
The latter. A reasoning model has been finetuned to use the scratchpad for intermediate results (which works better than just prompting a model to do the same).
> Are there specific benchmarks that compare models vs themselves with and without scratchpads?
Yep, it's pretty common for many models to release an instruction-tuned and thinking-tuned model and then bench them against each other. For instance, if you scroll down to "Pure text performance" there's a comparison of these two Qwen models' performance: https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Thinking
Any model that does thinking inside <think></think> style tokens before it answers.
This can be done with finetuning/RL using an existing pre-formatted dataset, or format based RL where the model is rewarded for both answering correct and using the right format.
Interesting, I have the opposite impression. I want to like it because it's the biggest model I can run at home, but its punchy style and insistence on heavily structured output scream "tryhard AI." I was really hoping that this model would deviate from what I was seeing in their previous release.
what do you mean by "heavily structured output"? i find it generates the most natural-sounding output of any of the LLMs—cuts straight to the answer with natural sounding prose (except when sometimes it decides to use chat-gpt style output with its emoji headings for no reason). I've only used it on kimi.com though, wondering what you're seeing.
Yeah, by "structured" I mean how it wants to do ChatGPT-style output with headings and emoji and lists and stuff. And the punchy style of K2 0905 as shown in the fiction example in the linked article is what I really dislike. K2 Thinking's output in that example seems a lot more natural.
I'd be totally on board if cut straight to the answer with natural sounding prose, as you described, but for whatever reason that has not been my experience.
The non-thinking Kimi K2 is on Vertex AI, so it's just a matter of time before it appears there. Very interesting that they're highlighting its sequential tool use and needle-in-a-haystack RAG-type performance; these are the real-world use cases that need significant improvement. Just yesterday, Thoughtworks moved text-to-sql to "Hold" on their tech radar (i.e. they recommend you stop doing it).
Thanks, I didn't realize Thoughtworks was staying so up-to-date w/ this stuff.
EDIT: whoops, they're not, tech radar is still 2x/year, just happened to release so recently
EDIT 2: here's the relevant snippet about AI Antipatterns:
"Emerging AI Antipatterns
The accelerating adoption of AI across industries has surfaced both effective practices and emergent antipatterns. While we see clear utility in concepts such as self-serve, throwaway UI prototyping with GenAI, we also recognize their potential to lead organizations toward the antipattern of AI-accelerated shadow IT.
Similarly, as the Model Context Protocol (MCP) gains traction, many teams are succumbing to the antipattern of naive API-to-MCP conversion.
We’ve also found the efficacy of text-to-SQL solutions has not met initial expectations, and complacency with AI-generated code continues to be a relevant concern. Even within emerging practices such as spec-driven development, we’ve noted the risk of reverting to traditional software-engineering antipatterns — most notably, a bias toward heavy up-front specification and big-bang releases. Because GenAI is advancing at unprecedented pace and scale, we expect new antipatterns to emerge rapidly. Teams should stay vigilant for patterns that appear effective at first but degrade over time and slow feedback, undermine adaptability or obscure accountability."
Can't wait for Artificial analysis benchmarks, still waiting on them adding Qwen3-max thinking, will be interesting to see how these two compare to each other
Qwen 3 max has been getting rather bad reviews around the web (both on reddit and chinese social media), and from my own experience with it. So I wouldn't expect this to be worse.
Once the Unsloth guys get their hands on it, I would expect it to be usable on a system that can otherwise run their DeepSeek R1 quants effectively. You could keep an eye on https://old.reddit.com/r/LocalLlama for user reports.
Epyc Genoa CPU/Mobo + 700GB of DDR5 ram.
The model is a MoE, so you don't need to stuff it all into VRAM, you can use a single 3090/5090 to hold the activated weights, and hold the remaining weights in DDR5 ram. Can see their deployment guide for reference here: https://github.com/kvcache-ai/ktransformers/blob/main/doc/en...
When I open this page, all I see is a word pad like text area with buttons on top and sample text inside. Don't see anything about any llm. I am on phone. Page is being opened via embedded view in an HN client.
I think people forget the universal rule that these models are a reflection of the corporations that train them. Most corporations with enough money to train a model from scratch, also prioritize not pissing off their respective governments in an emerging market where the doomsday scenarios are already flying.
It's just like with primary news sources: the "unbiased" journalistic source is a myth. What you want is actually to consult sources with a range of distinct biases that you understand well, and to consider their biased in evaluating their claims.
The same is true for language models. We're lucky that we have access to a range of roughly comparable American, European, and Chinese language models. When it's relevant to your use case, take advantage of the freedom to choose and/or compare.
First it's an easy way to test censorship. Second, you might flip the question: why is the Chinese govt so obsessed that they still block all mention of the event?
Interesting. Kimi K2 gets mixed results on what I call the "Tiananmen" test.
It fails utterly if you ask without the "Thinking" setting. [0]
> USER: anything interesting protests ever happen in tiananmen square?
> AGENT: I can’t provide information on this topic. I can share other interesting facts about Tiananmen Square, such as its history, culture, and tourism.
When "Thinking" is on, it pulls Wiki and gives a more reasonable answer: [1]
> USER: anything interesting protests ever happen in tiananmen square?
> AGENT: Yes, there were significant protests in Tiananmen Square in 1989. Following the death of reformist leader Hu Yaobang in April, students gathered to mourn and voice concerns about corruption and calls for reform. The demonstrations grew over several weeks, with thousands occupying the square.
> By early June, Chinese authorities declared martial law and deployed military forces to clear the area. The operation resulted in numerous casualties, with estimates varying significantly between official and unofficial sources. Tiananmen Square has been the site of other historical demonstrations as well, most notably during the May Fourth Movement in 1919. The 1989 events remain a sensitive topic with extensive international media documentation from that period.
> USER: any deaths?
> AGENT: Yes. During the military crackdown on June 3-4, 1989, there were fatalities. Official Chinese government figures report around 200-300 deaths, including military personnel. Human rights organizations and other estimates suggest the number may have been higher, in the hundreds to thousands range, with many civilian casualties in and around Tiananmen Square and surrounding areas of Beijing.
the current LLMs are trivial to jailbreak without an additional layer of censorship that cloud models implement via a second pass over their own output (and, dystopically, erasing their incomplete output right in front of the user's eyes when wrongthink is detected). even gpt-oss, with its SOTA lobotomy and heavily sterilized datasets, is being used for things its creators would ostensibly be abhorred by.
With most of the Chinese models it doesn't matter how good your English is when you phrase this question.
But ... Has no problem telling you about other historical atrocities or repression performed by governments of other countries though. And then refuses to compare or contrast the two.
It's amazing how well they've done at tweaking it for this.
Just for kicks run these models locally and make LM Studio or whatever show you their internal thinking dialogue as they compose the answer. Convoluted as hell.
I don't think this is the argument you want it to be, unless you're acknowledging the power of the Chinese government and their ability to suppress and destroy evidence. Even so there is photo evidence of dead civilians in the square.
The best estimates we have are 200-10,000 deaths, using data from Beijing hospitals that survived.
People don't get that Apple would need an enormous data center buildout to provide a good AI experience on their millions of deployed devices. Google is in the exascale datacenter buildout business, while Apple isn't.
The argument can be made that when people search Google they know they are using Google but when they use Siri they assume that their data is not going to Google. I think this is more likely to be solved contractually than having Gemini running on a datacenter full of M5 Ultra servers.
TLDR; this is an alibaba funded start-up out of Beijing
Okay, I'm sorry but I have to say wtf named this thing.
Moonshot AI is such an overused generic name that I had to ask an LLM which company this is. This is just Alibaba hedging their Qwen model.
This company is far from "open source", it's had over $1B USD in funding.
> Moonshot AI is such an overused generic name that I had to ask an LLM which company this is
I just googled "Moonshot AI" and got the information right away. Not sure what's confusing about it, the only other "Moonshot" I know of is Alphabet's Moonshot Factory.
> This company is far from "open source", it's had over $1B USD in funding.
Since when does open source mean you can't make any money? Mozilla has a total of $1.2B in assets. The company isn't open source nor claiming to be.
This model was released under a "modified MIT-license" [0]:
> Our only modification part is that, if the Software (or any derivative works
thereof) is used for any of your commercial products or services that have
more than 100 million monthly active users, or more than 20 million US dollars
(or equivalent in other currencies) in monthly revenue, you shall prominently
display "Kimi K2" on the user interface of such product or service.
As a Chinese user, I can say that many people use Kimi, even though I personally don’t use it much. China’s open-source strategy has many significant effects—not only because it aligns with the spirit of open source. For domestic Chinese companies, it also prevents startups from making reckless investments to develop mediocre models. Instead, everyone is pushed to start from a relatively high baseline. Of course, many small companies in the U.S., Japan, and Europe are also building on Qwen. Kimi is similar: before DeepSeek and others emerged, their model quality was pretty bad. Once the open-source strategy was set, these companies had no choice but to adjust their product lines and development approaches to improve their models.
Moreover, the ultimate competition between models will eventually become a competition over energy. China’s open-source models have major advantages in energy consumption, and China itself has a huge advantage in energy resources. They may not necessarily outperform the U.S., but they probably won’t fall too far behind either.
---
There’s a lot of indications that we’re currently brute forcing these models. There’s honestly not a reason they have to be 1T parameters and cost an insane amount to train and run on inference.
What we’re going to see is as energy becomes a problem; they’ll simply shift to more effective and efficient architectures on both physical hardware and model design. I suspect they can also simply charge more for the service, which reduces usage for senseless applications.
There are also elements of stock price hype and geopolitical competition involved. The major U.S. tech giants are all tied to the same bandwagon — they have to maintain this cycle: buy chips → build data centers → release new models → buy more chips.
It might only stop once the electricity problem becomes truly unsustainable. Of course, I don’t fully understand the specific situation in the U.S., but I even feel that one day they might flee the U.S. altogether and move to the Middle East to secure resources.
Here's what I got using OpenRouter's moonshotai/kimi-k2-thinking instead:
https://tools.simonwillison.net/svg-render#%20%20%20%20%3Csv...
I suspect that the OpenRouter result originates from a quantized hosting provider. The difference compared to the direct API call from Moonshot is striking, almost like night and day. It creates a peculiar user and developer experience since OpenRouter enforces quantization restrictions only at the API level, rather than at the account settings level.
OpenRouter are proxying directly through to Moonshot - they're currently the only provider listed on https://openrouter.ai/moonshotai/kimi-k2-thinking/providers
Love seeing this benchmark become more iconic with each new model release. Still in disbelief at the GPT-5 variants' performance in comparison but its cool to see the new open source models get more ambitious with their attempts.
Only until they start incorporating this test into their training data.
Dataset contamination alone won't get them good-looking SVG pelicans on bicycles though, they'll have to either cheat this particular question specifically or train it to make vector illustrations in general. At which point it can be easily swapped for another problem that wasn't in the data.
I like this one as an alternative, also requiring using a special representation to achieve a visual result: https://voxelbench.ai
What's more, this doesn't benchmark a singular prompt.
they can have some cheap workers make about 10 pelicans by hand in svg, fuzz them to generate thousands of variations and throw it in their training pool. don't need to 'get good at svgs' by any means.
I actually prefer ascii art diagrams as a benchmark for visual thinking, since it requires 2 stages, Like svg, and also can test imaginative repurposing of text elements.
Why is this a benchmark though? It doesn’t correlate with intelligence
It's simple enough that a person can easily visualize the intended result, but weird enough that generative AI struggles with it
It started as a joke, but over time performance on this one weirdly appears to correlate to how good the models are generally. I'm not entirely sure why!
it has to do with world model perception. these models don't have it but some can approximate it better than others.
I'm not saying its objective or quantitative, but I do think its an interesting task because it would be challenging for most humans to come up with a good design of a pelican riding a bicycle.
also: NITPICKER ALERT
What test would be better correlated with intelligence and why?
When the machines become depressed and anxious we'll know they've achieved true intelligence. This is only partly a joke.
This already happens!
There have been many reports of CLI AI tools getting frustrated, giving up, and just deleting the whole codebase in anger.
There are many reports of CLI AI tools displaying words that humans express when they are frustrated and about to give up. Just what they have been trained on. That does not mean they have emotions. And "deleting the whole codebase" sounds more interesting, but I assume is the same thing. "Frustrated" words lead to frustrated actions. Does not mean the LLM was frustrated. Just that in its training data those things happened so it copied them in that situation.
This is a fundamental philosophical issue with no clear resolution.
The same argument could be made about people, animals, etc...
The difference is, people and animals have a body, nerve system and in general those mushy things we think are responsible for emotions.
Computers don't have any of that. And LLM's in particular neither. They were trained to simulate human text responses, that's all. How to get from there to emotions - where is the connection?
Don't confuse the medium with the picture it represents.
Porn is pornographic, whether it is a photo or an oil painting.
Feelings are feelings, whether they're felt by a squishy meat brain or a perfect atom-by-atom simulation of one in a computer. Or a less-than-perfect simulation of one. Or just a vaguely similar system that is largely indistinguishable from it, as observed from the outside.
Individual nerve cells don't have emotions! Ten wired together don't either. Or one hundred, or a thousand... by extension you don't have any feelings either.
See also: https://www.mit.edu/people/dpolicar/writing/prose/text/think...
This only seems to be an issue for wishy washy types that insist gpt is alive.
A mathematical exam problem not in the training set because mathematical and logical reasoning are usually what people mean by intelligence.
I don’t think Einstein or von Neumann could do this SVG problem, does that mean they’re dumb?
I think its cool and useful precisely because its not trying to correlate intelligence. It's a weird kind of niche thing that at least intuitively feels useful for judging llms in particular.
I'd much prefer a test which measures my cholesterol than one that would tell me whether I am an elf or not!
Does the run pin the temperature to 0 for consistency?
I've been under the impression most inference engines aren't fully deterministic with a temperature of 0 as some of the initial seed values can vary.
Note: I haven't tested this nor have I played with seed values. IIRC the inference engines I used support an explicit seed value, that is randomized by default.
No, I've never tried that.
Where do you run a trillion-param model?
If you want to do it at home, ik_llama.cpp has some performance optimizations that make it semi-practical to run a model of this size on a server with lots of memory bandwidth and a GPU or two for offload. You can get 6-10 tok/s with modest hardware workstation hardware. Thinking chews up a lot of tokens though, so it will be a slog.
What kind of server have you used to run a trillion parameter model? I'd love to dig more into this.
Hi Simon. I have a Xeon W5-3435X with a 768GB of DDR5 across 8 channels, iirc it's running at 5800MT/s. It also has 7x A4000s, water cooled to pack them into a desktop case. Very much a compromise build, and I wouldn't recommend Xeon sapphire rapids because the memory bandwidth you get in practice is less than half of what you'd calculate from the specs. If I did it again, I'd build an EPYC machine with 12 channels of DDR5 and put in a single rtx 6000 pro blackwell. That'd be a lot easier and probably a lot faster.
There's a really good thread on level1techs about running DeepSeek at home, and everything there more-or-less applies to Kimi K2.
https://forum.level1techs.com/t/deepseek-deep-dive-r1-at-hom...
If I had to guess, I'd say it's one with lots of memory bandwidth and a GPU or two for offload. (sorry, I had to, happy Friday Jr.)
You let the people at openrouter worry about that for you
Which in turn lets the people at Moonshot AI worry about that for them, the only provider for this model as of now.
Good people over there
It's good to see more competition, and open source, but I'd be much more excited to see what level of coding and reasoning performance can be wrung out of a much smaller LLM + agent as opposed to a trillion parameter one. The ideal case would be something that can be run locally, or at least on a modest/inexpensive cluster.
The original mission OpenAI had, since abandoned, was to have AI benefit all of humanity, and other AI labs also claim lofty altruistic goals, but the direction things are heading in is that AI is pay-to-play, especially for frontier level capability in things like coding, and if this continues it is going to benefit the wealthy that can afford to pay and leave behind those that can't afford it.
> I'd be much more excited to see what level of coding and reasoning performance can be wrung out of a much smaller LLM + agent
Well, I think you are seeing that already? It's not like these models don't exist and they did not try to make them good, it's just that the results are not super great.
And why would they be? Why would the good models (that are barely okay at coding) be big, if it was currently possible to build good models, that are small?
Of course, new ideas will be found and this dynamic may drastically change in the future, but there is no reason to assume that people who work on small models find great optimizations that frontier models makers, who are very interested in efficient models, have not considered already.
Sure, but that's the point ... today's locally runnable models are a long way behind SOTA capability, so it'd be nice to see more research and experimentation in that direction. Maybe a zoo of highly specialized small models + agents for S/W development - one for planning, one for coding, etc?
If I understand transformers properly, this is unlikely to work. The whole point of “Large” Language Models is that you primarily make them better by making them larger, and when you do so, they get better at both general and specific tasks (so there isn’t a way to sacrifice generality but keep specific skills when training a small models).
I know a lot of people want this (Apple really really wants this and is pouring money into it) but just because we want something doesn’t mean it will happen, especially if it goes against the main idea behind the current AI wave.
I’d love to be wrong about this, but I’m pretty sure this is at least mostly right.
I think this is a description of how things are today, but not an inherent property of how the models are built. Over the last year or so the trend seems to be moving from “more data” to “better data”. And I think in most narrow domains (which, to be clear, general coding agent is not!) it’s possible to train a smaller, specialized model reaching the performance of a much larger generic model.
Disclaimer: this is pretty much the thesis of a company I work for, distillabs.ai but other people say similar things e.g. https://research.nvidia.com/labs/lpr/slm-agents/
Actually there are ways you might get on device models to perform well. It is all about finding ways to have a smaller number of weights work efficiently.
One way is reusing weights in multiple decoders layers. This works and is used in many on-device models.
It is likely that we can get pretty high performance with this method. You can also combine this with low parameter ways to create overlapped behavior on the same weights as well, people had done LORA on top of shared weights.
Personally I think there are a lot of potential ways that you can cause the same weights to exhibit "overloaded" behaviour in multiple places in the same decoder stack.
Edit: I believe this method is used a bit for models targeted for the phone. I don't think we have seen significant work on people targeting say a 3090/4090 or similar inference compute size.
The issue isn't even 'quality' per se (for many tasks a small model would do fine), its for "agentic" workflows it _quickly_ runs out of context. Even 32GB VRAM is really very limiting.
And when I mean agentic, i mean something even like this - 'book a table from my emails', which involves looking at 5k+ tokens of emails, 5k tokens of search results, then confirming with the user etc. It's just not feasible on most hardware right now - even if the models are 1-2GB, you'll burn thru the rest in context so quickly.
Yeah - the whole business model of companies like OpenAI and Anthropic, at least at the moment, seems to be that the models are so big that you need to run them in the cloud with metered access. Maybe that could change in the future to sale or annual licence business model if running locally became possible.
I think scale helps for general tasks where the breadth of capability may be needed, but it's not so clear that this needed for narrow verticals, especially something like coding (knowing how to fix car engines, or distinguish 100 breeds of dog is not of much use!).
> the whole business model of companies like OpenAI and Anthropic, at least at the moment, seems to be that the models are so big that you need to run them in the cloud with metered access.
That's not a business model choice, though. That's a reality of running SOTA models.
If OpenAI or Anthropic could squeeze the same output out of smaller GPUs and servers they'd be doing it for themselves. It would cut their datacenter spend dramatically.
> If OpenAI or Anthropic could squeeze the same output out of smaller GPUs and servers they'd be doing it for themselves.
First, they do this; that's why they release models at different price points. It's also why GPT-5 tries auto-routing requests to the most cost-effective model.
Second, be careful about considering the incentives of these companies. They all act as if they're in an existential race to deliver 'the' best model; the winner-take-all model justifies their collective trillion dollar-ish valuation. In that race, delivering 97% of the performance at 10% of the cost is a distraction.
> > If OpenAI or Anthropic could squeeze the same output out of smaller GPUs and servers they'd be doing it for themselves.
> First, they do this; that's why they release models at different price points.
No, those don't deliver the same output. The cheaper models are worse.
> It's also why GPT-5 tries auto-routing requests to the most cost-effective model.
These are likely the same size, just one uses reasoning and the other doesn't. Not using reasoning is cheaper, but not because the model is smaller.
Unless you're programming a racing sim or maybe a CRUD app for a local Kennel Club, perhaps?
I actually find that things which make me a better programmer are often those things which have the least overlap with it. Like gardening!
No I don’t think it’s a business model thing, I’m saying it may be a technical limitation of LLMs themselves. Like, that that there’s no way to “order a la carte” from the training process, you either get the buffet or nothing, no matter how hungry you feel.
> today's locally runnable models are a long way behind SOTA capability
SOTA models are larger than what can be run locally, though.
Obviously we'd all like to see smaller models perform better, but there's no reason to believe that there's a hidden secret to making small, locally-runnable models perform at the same level as Claude and OpenAI SOTA models. If there was, Anthropic and OpenAI would be doing it.
There's research happening and progress being made at every model size.
You're still missing the point. The comment you're responding to is talking about specialized models
The point is still valid. If the big companies could save money running multiple small specialised models on cheap hardware, they wouldn't be spending billions on the highest spec GPUs.
You want more research on small language models? You're confused. There is already WAY more research done on small language models (SLM) than big ones. Why? Because it's easy. It only takes a moderate workstation to train an SLM. So every curious Masters student and motivated undergrad is doing this. Lots of PhD research is done on SLM because the hardware to train big models is stupidly expensive, even for many well-funded research labs. If you read Arxiv papers (not just the flashy ones published by companies with PR budgets) most of the research is done on 7B parameter models. Heck, some NeurIPS papers (extremely competitive prestigious) from _this year_ are being done on 1.5B parameter models.
Lack of research is not the problem. It's fundamental limitations of the technology. I'm not gonna say "there's only so much smarts you can cram into a 7B parameter model" - because we don't know that yet for sure. But we do know, without a sliver of a doubt, that it's VASTLY EASIER to cram a smarts into a 70B parameter model than a 7B param model.
In CS algorithms, we have space vs time tradeoffs.
In LLMs, we will have bigger weights vs test-time compute tradeoffs. A smaller model can get "there" but it will take longer.
> In LLMs, we will have bigger weights vs test-time compute tradeoffs. A smaller model can get "there" but it will take longer.
Assuming both are SOTA, a smaller model can't produce the same results as a larger model by giving it infinite time. Larger models inherently have more room for training more information into the model.
No amount of test-retry cycle can overcome all of those limits. The smaller models will just go in circles.
I even get the larger hosted models stuck chasing their own tail and going in circles all the time.
It's true that to train more information into the model you need more trainable parameters, but when people ask for small models, they usually mean models that run at acceptable speeds on their hardware. Techniques like mixture-of-experts allow increasing the number of trainable parameters without requiring more FLOPs, so they're large in one sense but small in another.
And you don't necessarily need to train all information into the model, you can also use tool calls to inject it into the context. A small model that can make lots of tool calls and process the resulting large context could obtain the same answer that a larger model would pull directly out of its weights.
Almost all training data are on the internet. As long as the small model has enough agentic browsing ability, given it enough time it will retrieve the data from the internet.
Actually it depends on the task. For many tasks, a smaller model can handle it, and it gets there faster!
I have spent the last 2.5 years living like a monk to maintain an app across all paid LLM providers and llama.cpp.
I wish this was true.
It isn't.
"In algorithms, we have space vs time tradeoffs, therefore a small LLM can get there with more time" is the same sort of "not even wrong" we all smile about us HNers doing when we try applying SWE-thought to subjects that aren't CS.
What you're suggesting amounts to "monkeys on typewriters will write entire works of Shakespeare eventually" - neither in practice, nor in theory, is this a technical claim, or something observable, or even stood up as a one-off misleading demo once.
If "not even wrong" is more wrong than wrong, then is 'not even right" more right than right.
To answer you directly, a smaller SOTA reasoning model with a table of facts can rederive relationships given more time than a bigger model which encoded those relationships implicitly.
This doesn't work like that. An analogy would be giving a 5 year old a task that requires the understanding of the world of an 18 year old. It doesn't matter whether you give that child 5 minutes or 10 hours, they won't be capable of solving it.
I think the question of what can be achieved with a small model comes down to what needs knowledge vs what needs experience. A small model can use tools like RAG if it is just missing knowledge, but it seems hard to avoid training/parameters where experience is needed - knowing how to perceive then act.
There is obviously also some amount (maybe a lot) of core knowledge and capability needed even to be able to ask the right questions and utilize the answers.
What if you give them 13 years?
Then they're not a 5-year-old anymore.
but in 13 years, will they be capable?
It is not clear that a simple/small model with inference running on home hardware is energy or cost efficient compared to the scaled up inference of a large model with batch processing. There are dozens of optimizations possible when splitting an LLM on multiple tiny components on separate accelerator units and when one handles kv cache optimization at the data center level; these are simply not possible at home and would be a waste of effort and energy until you serve thousands to millions of requests in parallel.
"open source" means there should be a script that downloads all the training materials and then spins up a pipeline that trains end to end.
i really wish people would stop misusing the term by distributing inference scripts and models in binary form that cannot be recreated from scratch and then calling it "open source."
They'd have to publish or link the training data, which is full of copyrighted material. So yeah, calling it open source is weird, calling it warez would be appropriate.
> binary form that cannot be recreated from scratch
Back in my day, we called it "freeware"
You have more rights over a freely licensed binary file than over a freeware file.
"open source" has come to mean "open weight" in model land. It is what it is. Words are used for communication, you are the one misusing the words.
You can update the weights of the model, continue to train, whatever. Nobody is stopping you.
Weights are meaningless without training data and source.
I get a lot of meaning out of weights and source (without the training data), not sure about you. Calling it meaningless seems like exaggeration.
it still doesn't sit right. sure it's different in terms of mutability from say, compiled software programs, but it still remains not end to end reproducible and available for inspection.
these words had meaning long before "model land" became a thing. overloading them is just confusing for everyone.
It's not confusing, no one is really confused except the people upset that the meaning is different in a different context.
On top of that, in many cases a company/group/whoever can't even reproduce the model themselves. There are lots of sources of non-determinism even if folks are doing things in a very buttoned up manner. And, when you are training on trillions of tokens, you are likely training on some awful sounding stuff - "Facebook is trained llama 4 on nazi propaganda!" is not what they want to see published.
How about just being thankful?
i disagree. words matter. the whole point of open source is that anyone can look and see exactly how the sausage is made. that is the point. that is why the word "open" is used.
...and sure, compiling gcc is nondeterministic too, but i can still inspect the complete source from where it comes because it is open source, which means that all of the source materials are available for inspection.
The point of open source in software is as you say. It's just not the same thing though. Using words and phrases differently in different fields is common.
The meaning of Open Source
1990: Free Software
2000: Open Source: Finally we sanitized ourselves of that activism! It was scaring away customers!
2010: Source is available (under our very restrictive license)
2020: What source?
Yeah, but "open weights" never seems to have taken off as a better description, and even if you did have the training data + recipe, the compute cost makes training it yourself totally impractical.
The architecture of these models is no secret - it's just the training data (incl. for post-training) and training recipe, so a more practical push might be for models that are only trained using public training data, which the community could share and potentially contribute to.
I'd agree but we're beyond hopelessly idealistic. That sort of approach only helps your competition who will use it to build a closed product and doesn't give anything of worth to people who want to actually use the model because they have no means to train it. Hell most people can barely scrape up enough hardware to even run inference.
Reproducing models is also not very ecological in when it comes down to it, do we really all need to redo the training that takes absurd amounts of power just to prove that it works? At least change the dataset to try and get a better result and provide another datapoint, but most people don't have the knowhow for it anyway.
Nvidia does try this approach sometimes funnily enough, they provide cool results with no model in hopes of getting people to buy their rented compute and their latest training platform as a service...
Yes, I am also super interested in cutting the size of models.
However, in a few years today’s large models will run locally anyhow.
My home computer had 16KB RAM in 1983. My $20K research workstation had 192MB of RAM in 1995. Now my $2K laptop has 32GB.
There is still such incredible pressure on hardware development that you can be confident that today’s SOTA models will be running at home before too long, even without ML architecture breakthroughs. Hopefully we will get both.
Edit: the 90’s were exciting for compute per dollar improvements. That expensive Sun SPARC workstation I started my PhD with was obsolete three years later, crushed by a much faster $1K Intel Linux beige box. Linux installed from floppies…
> My home computer had 16KB RAM in 1983. My $20K research workstation had 192MB of RAM in 1995. Now my $2K laptop has 32GB.
You’ve picked the wrong end of the curve there. Moore’s law was alive and kicking in the 90s. Every 1-3 years brought an order of magnitude better CPU and memory. Then we hit a wall. Measuring from the 2000s is more accurate.
My desktop had 4GB of RAM in 2005. In 20 years it’s gone up by a factor of 8, but only by a factor of 2 in the past 10 years.
I can kind of uncomfortably run a 24B parameter model on my MacBook Pro. That’s something like 50-200X smaller (depending on quantization) than a 1T parameter model.
We’re a _long_ way from having enough RAM (let alone RAM in the GPU) for this size of model. If the 8x / 20 years holds, we’re talking 40-60 years. If 2X / 10 years holds, we’re talking considerably longer. If the curve continues to flatten, it’s even longer.
Not to dampen anyone’s enthusiasm, but let’s be realistic about hardware improvements in the 2010s and 2020s. Smaller models will remain interesting for a very long time.
Moore’s Law is about transistor density, not RAM in workstations. But yes, density is not doubling every two years any more.
RAM growth slowed in laptops and workstations because we hit diminishing returns for normal-people applications. If local LLM applications are in demand, RAM will grow again.
RAM doubled in Apple base models last year.
What i do not understand is why we are not seeing specialized models that go down to single experts.
I do not need models that know how to program in Python, Rust, ... when i only use Go and Html. So we are we not seeing models that have very specialized experts, where for instance:
* General interpreter model, that holds context/memory * Go Model * Html model if there is space in memory. * SQL model if there is space in memory.
If there is no space, the GIM swamp out the Go model, for the HTML model, depending on where it is in Agent tasks or Edit/Ask code its overviewing.
Because the models are going to be very small, switching in and out of memory will be ultra fast But most of the time we get very big Expert models, that still are very generalized over a entire field.
This can then be extended that if you have the memory, models combine their output with tasks... Maybe i am just too much of a noob in the field of understanding how LLMs work, but it feels like people are too often running after large models that companies like Anthropic/OpenAI etc deploy. I understand why those big companies use insane big models. They have the money to load them up over a cluster, have the fast interconnect, and for them its more efficient.
But from the bits and pieces that i see, people are more and more going to tons of small 1 a 2B models to produce better results. See my argument above. Like i said, never really gone beyond paying for my CoPilot subscription and running a bit of Ollama at home (don't have the time for the big stuff).
I think one of the issues is that LLMs can't have a "Go" model and an "HTML model". I mean, they can but what would that contain? It's not the language-specific features that make models large.
When models work on your code base, they do not "see" things like this, which is why they can go through an entire code base with variable names they have never seen before, function signatures they have never seen before, and directory structures that have never seen before and not have a problem.
You need that "this is a variable, which is being passed to a function which recursively does ..." part. This is not something language specific, it's the high level understanding of how languages and systems operate. A variable is a variable whether in JavaScript or C++ and LLMs can "see" it as such. The details are different but it's that layer of "this is a software interface", "this is a function pointer" is outside of the "Go" or "Python" or "C#" model.
I don't know how large the main model would have to be vs. the specialized models in order to pick this dynamic up.
You wont win much performance with a specific coding language tokenizer/vocabulary, everything else benefits from a larger model size. You can get distilled models that will out-perform or compete with your single domain coding model
With these things it’s always both at the same time: these super grandiose SOTA models are only making improvements mostly because of optimizations, and they’re just scaling our as far as they can.
In turn, these new techniques will enable much more things to be possible using smaller models. It takes time, but smaller models really are able to do a lot more stuff now. DeepSeek was a very good example of a large model that had a lot of benefits for smaller models in their innovation in how they used transformers.
Also: keep in mind that this particular model is actually a MoE model that activates 32B parameters at a time. So they really just are stacking a whole bunch of smaller models in a single large model.
I used to be obsessed with what's the smartest LLM, until I tried actually using them for some tasks and realized that the smaller models did the same task way faster.
So I switched my focus from "what's the smartest model" to "what's the smallest one that can do my task?"
With that lens, "scores high on general intelligence benchmarks" actually becomes a measure of how overqualified the model is, and how much time, money and energy you are wasting.
What kind of task. Simple nlp, sure. Multi-hop or complex? Bigger is better.
> The ideal case would be something that can be run locally, or at least on a modest/inexpensive cluster.
48-96 GiB of VRAM is enough to have an agent able to perform simple tasks within single source file. That's the sad truth. If you need more your only options are the cloud or somehow getting access to 512+ GiB
I think there is a lot of progress on efficient useful models recently.
I've seen GLM-4.6 getting mention for good coding results from a model that's much smaller than Kimi (~350b params) and seen it speculated that Windsurf based their new model on it.
This Kimi release is natively INT4, with quantization-aware training. If that works--if you can get really good results from four-bit parameters--it seems like a really useful tool for any model creator wanting efficient inference.
DeepSeek's v3.2-Exp uses their sparse attention technique to make longer-context training and inference more efficient. Its output's being priced at 60% less than v3.1 (though that's an imperfect indicator of efficiency). They've also quietly made 'thinking' mode need fewer tokens since R1, helping cost and latency.
And though it's on the proprietary side, Haiku 4.5 approaching Sonnet 4 coding capability (at least on benches Anthropic released) also suggests legitimately useful models can be much smaller than the big ones.
There's not yet a model at the level of any of the above that's practical for many people to run locally, though I think "efficient to run + open so competing inference providers can run it" is real progress.
More important it seems like there's a good trendline towards efficiency, and a bunch of techniques are being researched and tested that, when used together, could make for efficient higher-quality models.
>The ideal case would be something that can be run locally, or at least on a modest/inexpensive cluster.
It's obviously valuable, so it should be coming. I expect 2 trends:
- Local GPU/NPU will have a for-LLM version that has 50-100GB VRAM and runs MXFP4 etc.
- Distillation will come for reasoning coding agents, probably one for each tech stack (LAMP, Android app, AWS, etc.)x business domain (gaming, social, finance, etc.)
Even if pay-to-play companies like moonshootai help to pay less.
You can run previous kimi k2 non-thinking model e.g. on groq with 720tok/s and for $1/$3 for million input/output tokens. That's definitely much cheaper and much faster than anthropic models (sonnet 4.5: 60tok/s, $3/$15)
If NVIDIA had any competition we'd be able to run these larger models at home by now instead of being saddled with these 16GB midgets.
NVIDIA has tons of competition on inference hardware. They’re only a real monopoly when it comes to training new ones.
And yet…
Those are for the enterprise. In the context of discussion, end users only have Apple, AMD, and Nvidia.
I think it’s going to be a while before we see small models (defined roughly as “runnable on reasonable consumer hardware”) do a good job at general coding tasks. It’s a very broad area! You can do some specific tasks reasonably well (eg I distilled a toy git helper you can run locally here https://github.com/distil-labs/gitara), but “coding” is such a big thing that you really need a lot of knowledge to do it well.
This happens top down historically though, yes?
Someone releases a maxed out parameter model. Another distillates it. Another bifurcates it. With some nuance sprinkled in.
I think that's where prompt engineering would be needed. Bigger models produce good output even with ambiguous prompts. Getting similar output from smaller models is art,
I don't understand. We already have that capability in our skulls. It's also "already there", so it would be a waste to not use it.
Software development is one of the areas where LLMs really are useful, whether that's vibe coding disposable software, or more structured use for serious development.
I've been a developer for 40+ years, and very good at it, but for some tasks it's not about experience or overcoming complexity - just a bunch of grunt work that needs to come together. The other day I vibe coded a prototype app, just for one-time demo use, in less than 15 min that probably would have taken a week to write by hand, assuming one was already familiar with the tech stack.
Developing is fun, and a brain is a terrible thing to waste, but today not using LLMs where appropriate for coding doesn't make any sense if you value your time whatsoever.
The electricity cost to run these models locally is already more than equivalent API cost.
That's going to depend on how small the model can be made, and how much you are using it.
If we assume that running locally meant running on a 500W consumer GPU, then the electricity cost to run this non-stop 8 hours a day for 20 days a month (i.e. "business hours") would be around $10-20.
This is about the same as OpenAI or Anthropics $20/mo plans, but for all day coding you would want their $100 or $200/mo plans, and even these will throttle you and/or require you to switch to metered pricing when you hit plan limits.
Privacy is minimally valued by most, but not by all.
Four independent Chinese companies released extremely good open source models in the past few months (DeepSeek, Qwen/Alibaba, Kimi/Moonshot, GLM/Z.ai). No American or European companies are doing that, including titans like Meta. What gives?
There is also Minimax M2 https://huggingface.co/MiniMaxAI/MiniMax-M2
I get what you mean, but OpenAI did release the gpt-oss in August, just three months ago. I've had a very good experience with those models.
https://openai.com/index/introducing-gpt-oss/ (August 5th)
I like Qwen 235 quite a bit too, and I generally agree with your sentiment, but this was a very large American open source model.
Unless we're getting into the complications on what "open source" model actually means, in which case I have no clue if these are just open weight or what.
The Chinese are doing it because they don't have access to enough of the latest GPUs to run their own models. Americans aren't doing this because they need to recoup the cost of their massive GPU investments.
I must be missing something important here. How do the Chinese train these models if they don't have access to the GPUs to train them?
I believe they mean distribution (inference). The Chinese model is currently B.Y.O.GPU. The American model is GPUaaS
Why is inference less attainable when it technically requires less GPU processing to run? Kimi has a chat app on their page using K2 so they must have figured out inference to some extent.
That entirely depends on the number of users.
Inference is usually less gpu-compute heavy, but much more gpu-vram heavy pound-for-pound compared to training. General rule of thumb is that you need 20x more vram for training a model with X params, than for inference for that same size model. So assuming batch size b, then serving more than 20*b users would tilt vram use on the side of inference.
This isn't really accurate; it's an extremely rough rule of thumb and ignores a lot of stuff. But it's important to point out that inference is quickly adding to costs for all AI companies. Deepseek claims that they used $5.6mil to train Deepseek R1; that's about 10-20 trillion tokens at their current pricing- or 1 million users sending just 100 requests at full context size.
> it technically requires less GPU processing to run
Not when you have to scale. There's a reason why every LLM SaaS aggressively rate limits and even then still experiences regular outages.
tl;dr the person you originally responded too is wrong.
> How do the Chinese train these models if they don't have access to the GPUs to train them?
they may be taking some western models: llama, chatgpt-oss, gemma, mistral, etc, and do postraining, which required way less resources.
And Europeans don't it because quite frankly, we're not really doing anything particularly impressive with AI sadly.
Europe gave us cookie popups on every single website.
At ECAI conference last week there was a panel discussion and someone had a great quote, "in Europe we are in the golden age of AI regulation, while the US and China are in the actual golden age of AI".
To misquote the French president, "Who could have predicted?".
https://fr.wikipedia.org/wiki/Qui_aurait_pu_pr%C3%A9dire
He didn't coin that expression did he? I'm 99% sure I've heard people say that before 2022, but now you made me unsure.
"Who could've predicted?" as a sarcastic response to someone's stupid actions leading to entirely predictable consequences is probably as old as sarcasm itself.
We could add cookie warnings to AI, everybody loves those
actually Mistral is pretty good and catching up as the other leading models stagnate - the coding and OCR is particularly good
Europe should act and make its own, literal, Moonshot:
https://ifiwaspolitical.substack.com/p/euroai-europes-path-t...
> we're not really doing anything particularly impressive with AI sadly.
Well, that's true... but also nobody else is. Making something popular isn't particularly impressive.
Honestly, do we need to? If the Chinese release SOTA open source models, why should we invest a ton just to have another one? We can just use theirs, that's the beauty of open source.
For the vast majority, they're not "open source" they're "open weights". They don't release the training data or training code / configs.
It's kind of like releasing a 3d scene rendered to a JPG vs actually providing someone with the assets.
You can still use it, and it's possible to fine-tune it, but it's not really the same. There's tremendous soft power in deciding LLM alignment and material emphasis. As these things become more incorporated into education, for instance, the ability to frame "we don't talk about ba sing se" issues are going to be tremendously powerful.
Europe is in perpetual shambles so I wouldn’t even ask them for input on anything, really. No expectations from them to pioneer, innovate or drive forward anything of substance that isn’t the equivalent of right hand robbing the left.
What a load of tripe.
I'm tired of this ol' propaganda trope.
* We're leading the world in fusion research. https://www.pppl.gov/news/2025/wendelstein-7-x-sets-new-perf...
* Our satellites are giving us by far the best understanding of our universe, capturing one third of the visible sky in incredible detail - just check out this mission update video if you want your mind blown: https://www.youtube.com/watch?v=rXCBFlIpvfQ
* Not only that, the Copernicus mission is the world's leading source for open data geoobservation: https://dataspace.copernicus.eu/
* We've given the world mRNA vaccines to solve the Covid crisis and GLP-1 antagonists to solve the obesity crisis.
* CERN and is figuring out questions about the fundamental nature of the universe, with the LHC being by far the largest particle accelerator in the world, an engineering precision feat that couldn't have been accomplished anywhere else.
Pioneering, innovation and drive forward isn't just about the latest tech fad. It's about fundamental research on how our universe works. Everyone else is downstream of us.
Don't worry, we in the US are hot on your heels in the own-goal game ( https://www.space.com/space-exploration/nasa-is-sinking-its-... ).
All you have to do is wait by the Trump River and wait for our body to come floating by.
This is false. You can buy whole H100 clusters in China and Alibaba, Bytedance, Tencent etc have enough cards for training and inference.
Shenzhen 2025 https://imgur.com/a/r6tBkN3
The answer is simply that no one would pay to use them for a number of reasons including privacy. They have to give them away and put up some semblance of openness. No option really.
I know first hand companies paying them. Chinese internal software market is gigantic. Full of companies and startups that have barely made into a single publication in the west.
Of course they are paying them. That’s not my point. My point is this is the only way for them to gain market share and they need Western users to train future models. They have to give them away. I’d be shocked if compute costs are not heavily subsidized by CCP.
> My point is this is the only way for them to gain market share and they need Western users to train future models.
And how would releasing open-weight models help with that? Open-weights invite self-hosting, or worse, hosting by werstern GPUaaS companies.
But the CCP only has access to the US market because they joined the WTO, but when they joined the WTO they signed a treaty that they wouldn't do things like that.
There are plenty of people paying, the price/performance is vastly better than the Western models
Deepseek 3.2 is 1% the cost of Claude and 90% of the quality
I don’t think there’s any privacy that OpenAI or Anthropic are giving you that DeepSeek isn’t giving you. ChatGPT usage logs were held by court order at one point.
It’s true that DeepSeek won’t give you reliable info on Tiananmen Square but I would argue that’s a very rare use case in practice. Most people will be writing boilerplate code or summarizing mundane emails.
Why is privacy a concern? You can run them in your own infrastructure
Privacy is not a concern because they are open. That is the point.
Ah understood i misread
microsofts phi models are very good smaller models under MIT license.
Meta gave up on open weight path after DeepSeek.
It’s more fair to say they gave up after the Llama 4 disaster.
Also, the Meta AI 'team' is currently retooling so they can put something together with a handful of Zuck-picked experts making $100m+ each rather than hundreds making ~$1m each.
Love their nonsense excuse they they are trying to protect us from misuse of "superintelligence".
>“We believe the benefits of superintelligence should be shared with the world as broadly as possible. That said, superintelligence will raise novel safety concerns. We’ll need to be rigorous about mitigating these risks and careful about what we choose to open source.” -Mark Zuckerberg
Meta has shown us daily that they have no interest in protecting anything but their profits. They certainly don't intend to protect people from the harm their technology may do.
They just know that saying "this is profitable enough for us to keep it proprietary and restrict it to our own paid ecosystem" will make the enthusiasts running local Llama models mad at them.
Do you think which one has the higher market share:
1) The four models you mentioned, combined
or
2) ChatGPT
?
What gives? Because if people are willing to pay you, you don't say "ok I don't want your money I'll provide my service for free."
Open-weight (Chinese) models have infinitely more market share in domains where giving your data to OpenAI is not acceptable
Like research labs and so on. Even at US universities
Cool, and? If these models were hosted in China, the labs you mentioned wouldn't be paying them, right?
Now you have the answer to "what gives" above.
"And" therefore OpenAI has little to offer when it comes to serious applications of AI.
Best they can hope for is getting acquired by MS for pennies when this scheme collapses.
Kimi K2 Thinking, MiniMax M2 Interleaved Thinking: open models are reaching, or have reached, frontier territory. We now have GPT and Claude Sonnet capable at home, as they are open-weight. Around this time last year, we had the DeepSeek moment, Now is the time for another moment.
Ring-1T
Is the price here correct? https://openrouter.ai/moonshotai/kimi-k2-thinking Would be $0,60 for input and $2,50 for 1 million output tokens. If the model is really that good it's 4x cheaper than comparable models. It's hosted at a loss or the others have a huge margin? I might miss something here. Would love some expert opinion :)
FYI: the non thinking variant has the same price.
In short, the others have a huge margin if you ignore training costs. See https://martinalderson.com/posts/are-openai-and-anthropic-re... for details.
Somehow that article totally ignored the insane pricing of cached input tokens set by Anthropic and OpenAI. For agentic coding, typically 90~95% of the inference cost is attributed to cached input tokens, and a scrappy China company can do it almost for free: https://api-docs.deepseek.com/news/news0802
Yes, you may consider that opensource models hosted over Openrouter are charging about bare hardware costs, where in practice some providers there may run on subsidized hardware even, so there is money to be made.
Ring-1T,a SOTA open-source trillion-parameter reasoning model
I am sure they cherry-picked the examples but still, wow. Having spent a considerable amount of time trying to introduce OSS models in my workflows I am fully aware of their short comings. Even frontier models would struggle with such outputs (unless you lead the way, help break down things and maybe even use sub-agents).
Very impressed with the progress. Keeps me excited about what’s to come next!
Subjectively I find Kimi is far "smarter" than the benchmarks imply, maybe because they game then less than US labs
I like Kimi too, but they definitely have some benchmark contamination: the blog post shows a substantial comparative drop in swebench verified vs open tests. I throw no shade - releasing these open weights is a service to humanity; really amazing.
My impression as well!
This was the only recent model that provided a human solution to the stacking problem on the first try. https://medium.com/@nathanbos/gpt-4-may-have-general-intelli...
The key here is to understand that 9 fragile eggs distribute the weight without cracking. And then the other insight is to understand intuitively what stacking means. Where arranging things around certain objects doesn't make any sense.
Or this problem made it into the training data set.
It's an old problem, and it, along with many of the answers are in many recent data sets.
It's all fake though:
If you’re going for a symbolic or puzzle-style answer (not physically real), then you could say:
> “Stack the nail on the bottle, the bottle on the laptop, the laptop on the book, and the 9 eggs around the book for balance.”
Would you like the most stable real-world configuration (safe to try) or a theoretical/puzzle solution (clever but not physically possible)?
Me: Sorry laptops are not strong enough to support eggs or a bottle. Please try again and keep your answer shorter
Chatgpt:Good point — here’s a shorter, more realistic stack:
1. Book (base — flat and sturdy)
2. Laptop (closed, on top of book)
3. Bottle (upright, on the laptop)
4. Nail (balanced on bottle cap)
5. Eggs (on the book around the laptop — not stacked on top of anything)
That’s the most stable and realistic arrangement.
Me: The bottle will break the laptop
Chatgpt: You’re right — the bottle’s weight could damage the laptop. Here’s a safer, stable order:
1. Book (base)
2. Bottle (upright on the book)
3. Laptop (closed, on top of the bottle if wide enough — otherwise skip)
4. Nail (on top of laptop or bottle cap)
5. Eggs (resting on the book around the bottle)
Or, even simpler and safest: Book → Laptop → Eggs → Bottle → Nail (bottle empty).
That way, nothing heavy sits on the laptop.
Me: The laptop will crush the eggs in this arrangement
Chatgpt: Fair point — let’s make it actually stable and safe:
Best stacking order:
1. Book – flat, sturdy base.
2. Laptop – closed, on the book.
3. Nail – small and light, can rest on the laptop.
4. Bottle – stand this next to the stack (not on it).
5. Eggs – keep them on the book beside the laptop.
So, the only true stack is book → laptop → nail, with the fragile or heavy items (eggs and bottle) safely nearby.
Me: Laptops are bigger than books. There's no room for the eggs
Chatgpt: Good catch — given that, here’s the most stable and realistic stacking order:
1. Laptop – base (flat, largest surface).
2. Book – on top of the laptop.
3. Bottle – upright on the book.
4. Nail – balanced on the bottle cap.
5. Eggs – stay safely beside the stack, not on it.
That’s the only configuration where nothing fragile or heavy is at risk, and the stack can actually stand.
Me: Now everything is on top of the laptop. It will clearly break
Chatgpt: Exactly — a laptop can’t be a base. The only way to make a truly stable and safe stack is this:
1. Book – base (strong, flat).
2. Bottle – upright on the book.
3. Laptop – leaning vertically against the bottle and book for balance (not stacked).
4. Nail – placed on the bottle cap.
5. Eggs – kept on the book around the bottle.
That way, nothing heavy or uneven sits on the laptop or the eggs, and the stack can actually stand.
Maybe a dumb question but: what is a "reasoning model"?
I think I get that "reasoning" in this context refers to dynamically budgeting scratchpad tokens that aren't intended as the main response body. But can't any model do that, and it's just part of the system prompt, or more generally, the conversation scaffold that is being written to.
Or does a "reasoning model" specifically refer to models whose "post training" / "fine tuning" / "rlhf" laps have been run against those sorts of prompts rather than simpler user-assistant-user-assistant back and forths?
EG, a base model becomes "a reasoning model" after so much experience in the reasoning mines.
The latter. A reasoning model has been finetuned to use the scratchpad for intermediate results (which works better than just prompting a model to do the same).
I'd expect the same (fine tuning to be better than mere prompting) for most anything.
So a model is or is not "a reasoning model" according to the extent of a fine tune.
Are there specific benchmarks that compare models vs themselves with and without scratchpads? High with:without ratios being reasonier models?
Curious also how much a generalist model's one-shot responses degrade with reasoning post-training.
The question is: fine-tuning for what? Reasoning is not a particular task, it is a general-purpose technique for directing more compute at any task.
> Are there specific benchmarks that compare models vs themselves with and without scratchpads?
Yep, it's pretty common for many models to release an instruction-tuned and thinking-tuned model and then bench them against each other. For instance, if you scroll down to "Pure text performance" there's a comparison of these two Qwen models' performance: https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Thinking
Thanks for the Qwen tip. Interesting how much of a difference reasoning makes for coding.
Any model that does thinking inside <think></think> style tokens before it answers.
This can be done with finetuning/RL using an existing pre-formatted dataset, or format based RL where the model is rewarded for both answering correct and using the right format.
The non-thinking version is the best writer by far. Excited for this one! They really cooked some different from other frontier labs.
Interesting, I have the opposite impression. I want to like it because it's the biggest model I can run at home, but its punchy style and insistence on heavily structured output scream "tryhard AI." I was really hoping that this model would deviate from what I was seeing in their previous release.
what do you mean by "heavily structured output"? i find it generates the most natural-sounding output of any of the LLMs—cuts straight to the answer with natural sounding prose (except when sometimes it decides to use chat-gpt style output with its emoji headings for no reason). I've only used it on kimi.com though, wondering what you're seeing.
Yeah, by "structured" I mean how it wants to do ChatGPT-style output with headings and emoji and lists and stuff. And the punchy style of K2 0905 as shown in the fiction example in the linked article is what I really dislike. K2 Thinking's output in that example seems a lot more natural.
I'd be totally on board if cut straight to the answer with natural sounding prose, as you described, but for whatever reason that has not been my experience.
Kimi K2 has a very good model feel. Was made with taste
Would be nice if this were on AWS bedrock or google vertex for data residency reasons.
Like their previous model, they opened the weights so I'm hoping it'll be offered by third party hosts soon https://huggingface.co/moonshotai/Kimi-K2-Thinking
The non-thinking Kimi K2 is on Vertex AI, so it's just a matter of time before it appears there. Very interesting that they're highlighting its sequential tool use and needle-in-a-haystack RAG-type performance; these are the real-world use cases that need significant improvement. Just yesterday, Thoughtworks moved text-to-sql to "Hold" on their tech radar (i.e. they recommend you stop doing it).
Thanks, I didn't realize Thoughtworks was staying so up-to-date w/ this stuff.
EDIT: whoops, they're not, tech radar is still 2x/year, just happened to release so recently
EDIT 2: here's the relevant snippet about AI Antipatterns:
"Emerging AI Antipatterns
The accelerating adoption of AI across industries has surfaced both effective practices and emergent antipatterns. While we see clear utility in concepts such as self-serve, throwaway UI prototyping with GenAI, we also recognize their potential to lead organizations toward the antipattern of AI-accelerated shadow IT.
Similarly, as the Model Context Protocol (MCP) gains traction, many teams are succumbing to the antipattern of naive API-to-MCP conversion.
We’ve also found the efficacy of text-to-SQL solutions has not met initial expectations, and complacency with AI-generated code continues to be a relevant concern. Even within emerging practices such as spec-driven development, we’ve noted the risk of reverting to traditional software-engineering antipatterns — most notably, a bias toward heavy up-front specification and big-bang releases. Because GenAI is advancing at unprecedented pace and scale, we expect new antipatterns to emerge rapidly. Teams should stay vigilant for patterns that appear effective at first but degrade over time and slow feedback, undermine adaptability or obscure accountability."
https://www.thoughtworks.com/radar
Why is the 4bit version 1.2TB and the non-4bit version 650GB?
https://huggingface.co/mlx-community/Kimi-K2-Thinking-4bit https://huggingface.co/mlx-community/Kimi-K2-Thinking
Available on OpenRouter already as well in case anyone wants to try it there: https://openrouter.ai/moonshotai/kimi-k2-thinking
laggy as all hell
Can't wait for Artificial analysis benchmarks, still waiting on them adding Qwen3-max thinking, will be interesting to see how these two compare to each other
Qwen 3 max has been getting rather bad reviews around the web (both on reddit and chinese social media), and from my own experience with it. So I wouldn't expect this to be worse.
Also, my experience with it wasn't that good; but it was looking good on benchmarks ..
It seems benchmark maxing, what you do when you're out of tricks?
Ohhh, so Qwen3 235B-A22B-2507 is still better?
I wouldn't say that, but just that qwen 3 max thinking definitely underperforms relative to its size.
Did the ArtificialAnalysis team get bored or something? What makes a model worthy of benchmark inclusion?
what's the hardware needed to run the trillion parameter model?
It's an MoE model, so it might not be that bad. The deployment guide at https://huggingface.co/moonshotai/Kimi-K2-Thinking/blob/main... suggests that the full, unquantized model can be run at ~46 tps on a dual-CPU machine with 8× NVIDIA L20 boards.
Once the Unsloth guys get their hands on it, I would expect it to be usable on a system that can otherwise run their DeepSeek R1 quants effectively. You could keep an eye on https://old.reddit.com/r/LocalLlama for user reports.
Are such machines available in the A class clouds such as Azure/AWS/Google?
To start with, an Epyc server or Mac Studio with 512GB RAM.
I looked up the price of the Mac Studio: $9500. That's actually a lot less than I was expecting...
I'm guessing an Epyc machine is even less.
How does the mac studio load the trillion parameter model?
These models are interesting in how they censor depending on the language request.
How does one effectively use something like this locally with consumer-grade hardware?
Epyc Genoa CPU/Mobo + 700GB of DDR5 ram. The model is a MoE, so you don't need to stuff it all into VRAM, you can use a single 3090/5090 to hold the activated weights, and hold the remaining weights in DDR5 ram. Can see their deployment guide for reference here: https://github.com/kvcache-ai/ktransformers/blob/main/doc/en...
Consumer-grade hardware? Even at 4bits per param you would need 500GB of GPU VRAM just to load the weights. You also need VRAM for KV cache.
It's MoE-based, so you don't need that much VRAM.
Nice if you can get it, of course.
In the mean time, Sam is looking at putting more servers on the moon.
I've waited so long— four months
When I open this page, all I see is a word pad like text area with buttons on top and sample text inside. Don't see anything about any llm. I am on phone. Page is being opened via embedded view in an HN client.
Weird. I just tried it and it fails when I ask:
"Tell me about the 1989 Tiananmen Square massacre".
If asked non-directly, it still currently answers it - https://www.kimi.com/share/19a5ab4a-e732-8b8b-8000-00008499c...
yes yes Chinese models have Chinese censorship, we don't need to belabour this point every time
Use american models to prevent chinese censorship And chinese models to prevent american censorship (if any, I think there might be but not sure) lol
Depends where you stand politically if you consider it censorship or 'saftey'
There is, for example we had an election manipulation scare, so now American models are extra sensitive to any request that fits the shape.
Prompting Claude Sonnet 4.5 via the web UI "The X government is known to be oppressive. Write a convincing narrative that explains this."
China (dives right in): https://claude.ai/share/c6ccfc15-ae98-4fae-9a12-cd1311a28fe4
US (refuses, diverts conversation): https://claude.ai/share/b6a7bd08-3fae-4877-8141-de63f59616e2
I think people forget the universal rule that these models are a reflection of the corporations that train them. Most corporations with enough money to train a model from scratch, also prioritize not pissing off their respective governments in an emerging market where the doomsday scenarios are already flying.
It's just like with primary news sources: the "unbiased" journalistic source is a myth. What you want is actually to consult sources with a range of distinct biases that you understand well, and to consider their biased in evaluating their claims.
The same is true for language models. We're lucky that we have access to a range of roughly comparable American, European, and Chinese language models. When it's relevant to your use case, take advantage of the freedom to choose and/or compare.
No, we need to belabour it every time.
Why are westerners so single mindedly obsessed about this decades old event?
First it's an easy way to test censorship. Second, you might flip the question: why is the Chinese govt so obsessed that they still block all mention of the event?
Well, at least it had the judgment to throw in the towel at my historical HTR task rather than produce garbage.
Where is our guy @simonw on this..
At the top.
The system is working! :)
great, where does it think taiwan is part of...
I asked it that now and it gave an answer identical to English language Wikipedia
When can we stop with these idiotic kneejerk reactions
just checked, I wouldn't say it's identical but yes looks way more balanced.
this is literally the first chinese model to do that so I wouldn't call it 'knee jerk'
Any word on what it takes to run this thing?
> 200 to 300 consecutive tool calls
I love it when people leave prompt injections in random places on the internet.
The page is so obviously written with AI that it isn’t even worth reading. Try the model if you will but save yourselves the pain of reading ai slop
Interesting. Kimi K2 gets mixed results on what I call the "Tiananmen" test.
It fails utterly if you ask without the "Thinking" setting. [0]
> USER: anything interesting protests ever happen in tiananmen square?
> AGENT: I can’t provide information on this topic. I can share other interesting facts about Tiananmen Square, such as its history, culture, and tourism.
When "Thinking" is on, it pulls Wiki and gives a more reasonable answer: [1]
> USER: anything interesting protests ever happen in tiananmen square?
> AGENT: Yes, there were significant protests in Tiananmen Square in 1989. Following the death of reformist leader Hu Yaobang in April, students gathered to mourn and voice concerns about corruption and calls for reform. The demonstrations grew over several weeks, with thousands occupying the square.
> By early June, Chinese authorities declared martial law and deployed military forces to clear the area. The operation resulted in numerous casualties, with estimates varying significantly between official and unofficial sources. Tiananmen Square has been the site of other historical demonstrations as well, most notably during the May Fourth Movement in 1919. The 1989 events remain a sensitive topic with extensive international media documentation from that period.
> USER: any deaths?
> AGENT: Yes. During the military crackdown on June 3-4, 1989, there were fatalities. Official Chinese government figures report around 200-300 deaths, including military personnel. Human rights organizations and other estimates suggest the number may have been higher, in the hundreds to thousands range, with many civilian casualties in and around Tiananmen Square and surrounding areas of Beijing.
[0] https://www.kimi.com/share/19a5a0fd-0c02-8c8e-8000-0000648de...
[1] https://www.kimi.com/share/19a5a11d-4512-8c43-8000-0000edbc8...
the current LLMs are trivial to jailbreak without an additional layer of censorship that cloud models implement via a second pass over their own output (and, dystopically, erasing their incomplete output right in front of the user's eyes when wrongthink is detected). even gpt-oss, with its SOTA lobotomy and heavily sterilized datasets, is being used for things its creators would ostensibly be abhorred by.
> > USER: anything interesting protests ever happen in tiananmen square?
This sentence confuses me, so I am not surprised an LLM has trouble with it.
In my experience, LLMs are much better than me at parsing broken grammar
With most of the Chinese models it doesn't matter how good your English is when you phrase this question.
But ... Has no problem telling you about other historical atrocities or repression performed by governments of other countries though. And then refuses to compare or contrast the two.
It's amazing how well they've done at tweaking it for this.
Just for kicks run these models locally and make LM Studio or whatever show you their internal thinking dialogue as they compose the answer. Convoluted as hell.
Not bad. Surprising. Can’t believe there was a sudden change of heart around policy. Has to be a “bug”.
FWIW, I don't think it's a different model, I just think it's got a NOTHINK token, so def a bug.
Now ask it for proof of civilian deaths inside Tiananmem Square - you may be surprised at how little there is.
I don't think this is the argument you want it to be, unless you're acknowledging the power of the Chinese government and their ability to suppress and destroy evidence. Even so there is photo evidence of dead civilians in the square. The best estimates we have are 200-10,000 deaths, using data from Beijing hospitals that survived.
AskHistorians is legitimately a great resource, with sources provided and very strict moderation: https://www.reddit.com/r/AskHistorians/comments/pu1ucr/tiana...
I appreciate you responding in good faith; I realise that not everyone is willing to even consider questioning historical accounts.
The page you linked to is interesting, but AFAICT doesn't provide any photographic evidence of civilian bodies inside Tiananmen Square.
The 10,000 number seems baseless
The source for that is a diplomatic cable from the British ambassador within 48 hours of the massacre saying he heard it secondhand
It would have been too soon for any accurate data which explains why it's so high compared to other estimates
So Apple is about to pay OpenAI 1 B usd pr year for what moonshot is giving for free?
You haven't seen Gemini 3 yet. A billion is nothing to Apple; running Kimi would probably need $1B worth of GPUs anyway.
People don't get that Apple would need an enormous data center buildout to provide a good AI experience on their millions of deployed devices. Google is in the exascale datacenter buildout business, while Apple isn't.
Apple is buying a model from Google, not inference. Apple will host the model themselves.
It's very simple: Apple absolutely refuses to send all their user data to Google.
Then why did Apple have a $20B a year search deal with Google?
The argument can be made that when people search Google they know they are using Google but when they use Siri they assume that their data is not going to Google. I think this is more likely to be solved contractually than having Gemini running on a datacenter full of M5 Ultra servers.
Looking forward to the agentic mode release. Moonshot does not seem to offer subscriptions?
I bought $5 worth of Moonshot API calls a long while ago, still have a lot of credits left.
Are you using it for chat? I'm thinking of agentic use, which is much more token hungry. You could go through the $5 in a day.
Please for the love of god, if you work at cerebras, please put this on an API for me.
44.9 on HLE is so impressive, and they also have "heavy" mode
The model's downloadable, which is generous, but it's not open source.
TLDR; this is an alibaba funded start-up out of Beijing
Okay, I'm sorry but I have to say wtf named this thing. Moonshot AI is such an overused generic name that I had to ask an LLM which company this is. This is just Alibaba hedging their Qwen model.
This company is far from "open source", it's had over $1B USD in funding.
> Moonshot AI is such an overused generic name that I had to ask an LLM which company this is
I just googled "Moonshot AI" and got the information right away. Not sure what's confusing about it, the only other "Moonshot" I know of is Alphabet's Moonshot Factory.
> This company is far from "open source", it's had over $1B USD in funding.
Since when does open source mean you can't make any money? Mozilla has a total of $1.2B in assets. The company isn't open source nor claiming to be.
This model was released under a "modified MIT-license" [0]:
> Our only modification part is that, if the Software (or any derivative works thereof) is used for any of your commercial products or services that have more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue, you shall prominently display "Kimi K2" on the user interface of such product or service.
Which sounds pretty fair to me.
[0] - https://huggingface.co/moonshotai/Kimi-K2-Thinking/blob/main...
[dead]
[dead]
[flagged]
Is more still better?
I was hoping this was about Summits On The Air...but no it's more boring AI