Open source AI is actually a lot less "powerful" than genuine frontier models, i.e. it has a much tighter inherent capability ceiling. This is "decelerationist" from a purely AGI-pilled point of view but is actually great if you're worried about a capabilities arms race putting AI Safety at severe risk.
Kimi K3 is plausibly a lot less dangerous than a totally jailbroken ChatGPT/Gemini/Claude Sonnet (let alone Opus or Fable!) and it's quite deeply weird how no one seems to be calling for those models to be banned or restrained by further regulation. Why the double standard against the less concerning (but more efficient!) open weight models?
Open Source models decelerate growth of closed AI. For people who think (or want) AI = closed_AI then that argument has weight. Good luck getting them to update their priors.
> 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
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.
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.
Interesting license. They basically added a "no composer / better models" if revenues (including downstream) above 20m clause, where you'd have to get a separate agreement with them.
Still has the attribution clause if serving this to your users.
> we build a self-evolving, hierarchically organized knowledge graph that agents continuously expand through web-scale exploration across knowledge-intensive and coding domains
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).
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?
Latest "big" release from any of the bigger American lab must have been GPT-OSS-120b I think? Released ~summer 2026, so pretty much a year ago. Doesn't seem like it'll happen by itself, so something either forcing their hand figuratively, or something forcing their hand literally.
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.
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
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?)
Kimi K3 is plausibly a lot less dangerous than a totally jailbroken ChatGPT/Gemini/Claude Sonnet (let alone Opus or Fable!) and it's quite deeply weird how no one seems to be calling for those models to be banned or restrained by further regulation. Why the double standard against the less concerning (but more efficient!) open weight models?
I think you meant less research and experiments in big labs because they don't get all the AI money.
Training is expensive, but they also have more than 10 000 of employees combined and they cost a lot of money.
> 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
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
What was it?
1 - https://x.com/Kimi_Moonshot/status/2035074972943831491?lang=...
Still has the attribution clause if serving this to your users.
Other than that, free to do whatever you want.
That's interesting!
maybe some interesting theoretical work there around the rate of production of new knowledge itself and various mechanisms (human approaches, mechanistic approaches, etc).
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.
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.