Kimi-K3 Technical Report [pdf]

2 hours ago (github.com)

Related: Kimi-K3 on HuggingFace - https://news.ycombinator.com/item?id=49065752

Back of the envelope calculation (could be off, correct me if I am)

If you are a large enough company that spends million+ on inference a month, it makes sense to buy a GB300 rack ($6M on top range from what I could find) which has 20.7 TB. Since the model is mixed trained (MXFP4), you would need less than 10% of the rack's memory to serve the full model. Aggregate HBM bandwidth: 576 TB/s. You can run over 6000 parallel agentic workflows (each with ~100k context on average) at ~30 tok/s.

Assuming the annual amortization+electricity at $1.5M/year and about 50% average annual utilization, you get less than 60 cents (USD) per million output token, for a frontier model with plenty of capacity to share, all your data never leaving premises and well over an order of magnitude cheaper!

As long as a company believes that the openweight models will continue to get more capable and 'AI is here to stay', this model provides the first solid footing for a decision to just buy a rack.

  • And hire 2 or 3 dev ops to keep it running ?

    That another 400 to 700k.

    It becomes your problem and not someone else’s. However, I don’t trust hosted LLMs for anything that needs to be private.

    • There will be cloud/SaaS vendors who have lower cost of labor/capital due to automation and financing terms.

      Having these models in the open caps the inference margin.

    • You’ll slap some training on existing technologists/infra/sysadmin folks and perhaps have a support contract for the edge cases (hardware troubleshooting and advanced replacement).

      (managed an entire data center building with thousands of servers a lifetime ago with ~2-3 other people, it’s only gotten easier over the last two decades imho)

  • I think the licensing that would likely apply to a company that's able to afford ~$6M rack and the associated infrastructure muddies this somewhat

    • I think internal use is allowed at any scale in the license?

      > 4. The requirements set forth in Sections 2 and 3 do not apply to: (a) internal use of the Software, defined as any use that does not make the Software, its outputs, or its underlying capabilities available to third parties; [...]

      1 reply →

    • As far as I can tell, license fees are only applicable if you have more than 20m USD/year revenue from services provided using the model, or serve more than 100m users.

Also open sourced a bunch of infra to go with it.

Anyone who claims open source and open weights models are "decel" needs to get their head checked

https://github.com/MoonshotAI/MoonEP

https://github.com/kvcache-ai/AgentEnv

https://github.com/MoonshotAI/FlashKDA

  • This comment would be much better without the second line

    • I'm not sure I understand the case for open-source models being decelerationist, is this it?

      Decel:

      - Potentially reduces investor appetite for funding big labs.

      - More risk of powerful AI getting in bad hands -> more regulation.

      Accel:

      - More competition so big labs can't rest on laurels.

      - More research in open, so all labs can accrete advancements faster.

      I feel like open-source = acceleration has a much more clear argument. (and how bad would deceleration be in any case?)

      10 replies →

    • Why? Absolutely correct, especially considering the position of the person they're referring to

    • I know it's a hard ask on this site, but I need you to start parsing content and not tone. It was a helpful bit of context, even if it was a bit vitriolic.

      7 replies →

  • To me its clear that it is decel

    the only reason other labs can catch up is because the frontier labs can be distilled, and they siphon a % of the labs' revenue to reinvest into the next iteration

    full accel would mean nationalizing the big 2 labs and locking in manhattan project style until RSI

    • only if you only get your news from main stream business press and Big Lab propaganda channels

      There's no chance K3 is a distill of Fable, it came out way too soon after the limited fable release to be feasbile.

      If you look at all of the top ML conferences, chinese labs contribute way more to advances in ML than "Open"AI and Anthropic: https://www.reddit.com/r/TheMachineGod/comments/1pi4q7f/pape...

      This K3 release just helped every other lab on the planet stay in the race by making it possible for them to build on top of it, placing them at the frontier starting line instead of having to spend billions of their own dollars and risking it all to attempt to catch up.

      The open source contributions I linked to above will move the whole field forward and reduce the costs of training and inference for everyone.

      Open science compounds on it self, every new advancement pushes the field forwards and opens up new grounds for future improvements.

      It is impossible for a single closed lab to consistently stay ahead of the rest of the field, especially in a huge growing research area like machine learning. The only only advantage the big labs have is money, but the naive scaling game is not sustainable long term when you have to pay 10-100x more then the fast followers and we start getting more and more open models or use case specific models that can handle 90% of high volume use cases.

      Research is a high variance, low expected value activity, meaning that the few large concentrated labs have to be conservative with their bets and double down on proven things when scaling up. The rest of the field is like a diversified portfolio, with thousands of players making smaller riskier bets that only require a few of them to succeed (like K3 did here, and DeepSeek a year ago)

      EDIT: also if you look at most of the work from OpenAI, it's mostly taking existing promising open research work and scaling it up. (except for things like CLIP and etc from Alec Radford)

      2 replies →

    • OpenAI's head of strategic futures publicly stated that you can't explain the quality of the newest Kimi via distillation.

      Further, you can just read the papers released alongside most open models. Plenty of hugely influential research results published that drive the frontier forward. It's not like these models are just existing architectures downloaded from Huggingface and trained on frontier lab APIs.

    • being able to replicate it in the open means that there's nothing special about frontier models.

      Frontier models would have to do something extraordinary or unique, or unreplicatable, because clearly there is no moat, and US companies are sitting on huge nvidia valuations and get surprised when competitors beat them.

License: https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE

> If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose.

+ the existing 100 million monthly active users, or more than 20 million US dollars for commercial products have to name Kimi clause

  • Kimi 2.6 had the same janky license: https://huggingface.co/moonshotai/Kimi-K2.6/blob/main/LICENS... - looks like they've been doing that at least as far back as K2.

    The models they released in 2025 - https://huggingface.co/moonshotai/models - were clean MIT. They started doing the "modified MIT" thing in January 2026 with moonshotai/Kimi-K2-Thinking

    • To be clear, the restriction on "Model as a Service" past $20M yearly revenue is new to K3. K2.x had the attribution requirement for any commercial use with more than $20M monthly revenue.

      1 reply →

    • In what way was the license "janky"? Moonshot isn't a trillion dollar US behemoth. If these terms allow it to release near frontier models with open weights, so startups and anyone with enough hardware can use them, while charging only companies with more than $20 million in revenue or 100 million users, that seems like a reasonable trade off.

      4 replies →

  • I wonder how they'll figure out who to target for litigation when this license is violated.

    • They have certainly trained in some secret call/response pairs that would uniquely identify Kimi serving.

    • How many companies host and serve models via API and have a $20M+ revenue? Going to be pretty straightforward to catch offenders.

    • Software licenses aren't enforced through litigation as much as they are enforced through the _threat_ of litigation and legal risk. In other words, pretty much every company pays lawyers to minimize legal risk. Those lawyers inevitably look at all the contracts, agreements, and software licenses, and tell the C-suite what to do in order to keep their legal exposure as low as possible. "Don't violate other companies' IP," is pretty low-hanging fruit in those conversations.

      It is a very rare (and ballsy, and perhaps incompetent) company that ignores their lawyers' recommendations to adhere to the letter of all of the software licenses they are bound to.

  • Hah, so much for "open weights" :D Fair enough, they're at least downloadable, shame they didn't end up being actually open, nor open source, the community had really high hopes for this. But again, still available for download, so better than nothing else I suppose.

> we build a self-evolving, hierarchically organized knowledge graph that agents continuously expand through web-scale exploration across knowledge-intensive and coding domains

That's interesting!

It's funny that we've finally returned to tanh activation functions, time is a circle.

  • From the paper (page 6 with a comparison to GLU and SwiGLU) they are not using tanh directly (i.e. f(x) = tanh(x)) but:

        f_gate(b,x) = b * tanh(x / b) * sigmoid(x)
    
        f_up(b,x) = b * tanh(x / b)
    

    Looking at the graph I wonder if this is to try and get the best of both GLU (better representation at higher values of x >~ 5) and SwiGLU (the value bump just before 0).

knowledge graph guided task synthesis. very cool! i have long wondered about the "how do you get good coverage of all the tasks" problem.

maybe some interesting theoretical work there around the rate of production of new knowledge itself and various mechanisms (human approaches, mechanistic approaches, etc).

What would be the current best method to fine-tune it for my own specific agentic tasks? LoRA + DPO? GRPO? Something else?

  • LoRA + SFT, but it'll be big - better to wait for a finetuning API from one of the providers, I wouldn't jump straight to RL or off-policy pseudo-RL like DPO.

Can someone explain what are teachers in Multi-Teacher On-Policy Distillation? I can imagine math, coding and other verifiable domains, but they also have biology? Is it where distillation from bigger models come in?

  • They reference https://thinkingmachines.ai/blog/on-policy-distillation/

    If I understand correctly, it's distillation via having a teacher model score each of the student's tokens for a problem based on their own probabilities of generating that token at each step in the sequence. The reward/loss is then applied as RL.

    The multi-teacher bit seems to imply they're distilling from multiple models. It's light on the details, but it seems like it could be part of distilling from frontier/closed models. Provided they calculate the logprobs, which OpenAI seems to allow via API but not Anthropic. Maybe they have a way of estimating the logprobs externally?

    This method can be used to learn any domain from the teacher. Biology included.

    • It's not that light on the details. I read the paper and they train several different models in parallel over a few different domains and then they distill from their own models to get the final model.

Does anyone know if a torrent is available? I think it would take quite a while to download 1.5tb from their servers.

This is amazing to witness. Moonshot open sourcing Kimi K3, a frontier AI and other components really means we are getting abundant AI for all of humanity.

Kudos to Moonshot for truly being what OpenAI should have been.

Fable-level and frontier AI should be open source and available to everyone for free.

  • > Kudos to Moonshot for truly being what OpenAI should have been.

    Kudos to Moonshot for making these weights available for download. Lets not fool ourselves and claim these are "open source" by any understanding of the concept though, there are usage restrictions (even if you download them) and also training data isn't clearly broken down either, nor it it actually using a FOSS license.

    • > Kudos to Moonshot for making these weights available for download. Lets not fool ourselves and claim these are "open source" by any understanding of the concept though, there are usage restrictions (even if you download them) and also training data isn't clearly broken down either, nor it it actually using a FOSS license.

      They will probably never release the training data because that represents a large part of their competitive moat. The same is true of US companies (Google, OpenAI, Meta etc) none of which has released the full training data for its open models.

      They use private datasets that cost a lot to acquire, synthetic datasets and a lot copyrighted material for which they don't have licensing.

      What matters most is that, with the necessary hardware, I can download a near frontier model, run it and modify it however I want. The other concerns you mentioned are just noise. And if my company is generating $20 million in revenue or serving 100 million users, it can probably afford a relatively inexpensive commercial license.

      4 replies →

What would it take to get an American open model to compete with this?

  • Latest "big" release from any of the bigger American lab must have been GPT-OSS-120b I think? Released ~summer 2025, so pretty much one years ago. Doesn't seem like it'll happen by itself, so something either forcing their hand figuratively, or something forcing their hand literally.

    Personally I was wishing/hoping for one of the recent Gemma releases to be in the ~100B class at least, but sadly Google is keeping that all for themselves.

  • if it's to be believed, it's so easy. you distill it, and you have a copy in 2 weeks, isn't that what China is doing? so 2 weeks from now, we should have one.

I would want to see three things before drawing strong conclusions:

End-to-end tokens/sec and cost on realistic coding agent trajectories, including tool outputs and retries, not isolated decode benchmarks.

Cache hit rates and prefill cost for branching, multi-turn sessions.

Router-load distributions after post-training, where expert collapse or specialization problems often show up.