Moonshot AI has kept its promise. On November 4, the Chinese lab published the full weights of Kimi K3.1 Pro on Hugging Face, twelve days after launching the model on October 23. Anyone can now download, fine-tune, and run one of the strongest coding agent models in the world, a model that Moonshot’s own tests put within a point of Anthropic’s Claude Opus 5.5.
The release matters because of its scale. Kimi K3.1 Pro is a 2.8 trillion parameter model, the largest open weight release on Hugging Face to date, ahead of Qwen3.8 at 2.4 trillion and DeepSeek V4 Pro at 1.6 trillion. It only activates 104 billion parameters per token, routed through 896 experts with 16 selected for each token, which is why it can be this large without needing a data center for every query. It also keeps the 1 million token context window from the K3 generation, and it accepts images as well as text.
The download is 1.56TB, and that is the point
Let us be honest about what “open weights” means here. The checkpoint ships as 96 files totaling about 1.56TB, stored at 4 bit precision after quantization aware training. That is far smaller than the roughly 5.6TB a 16 bit copy would need, but it is still a serious download that needs serious hardware. Moonshot recommends running it with vLLM, SGLang, or TokenSpeed, and most readers will never run it at home.
That does not make the release symbolic. Open weights mean independent researchers can test Moonshot’s benchmark claims instead of taking them on trust, startups can fine-tune the model for their own products without paying API bills, and rival labs can study exactly how it was built. The hosted API remains the easy route for most people, and it got cheaper too, at $2 per million input tokens and $10 per million output tokens, down a third from K3’s $3 and $15. A faster Highspeed variant costs $4 and $20.
How Kimi K3.1 Pro compares on benchmarks
Moonshot’s headline number is Terminal-Bench 4.0, a test of coding agents on real terminal tasks. Kimi K3.1 Pro scores 65.7 percent, against 66.4 percent for Claude Opus 5.5 and 58.2 percent for OpenAI’s GPT-6 Astra. Those are Moonshot’s own figures, not independent results, and the open weights now let third parties verify them. The independent Artificial Analysis Intelligence Index already ranks the model fourth overall at 55.1, up 11.5 points from K3.
| Model | Terminal-Bench 4.0 | API price per million tokens | Weights |
|---|---|---|---|
| Kimi K3.1 Pro | 65.7% | $2 input / $10 output | Open, published Nov 4 |
| Claude Opus 5.5 | 66.4% | Higher, closed model | Closed |
| GPT-6 Astra | 58.2% | Higher, closed model | Closed |
| Kimi K3 | Lower than K3.1 Pro | $3 input / $15 output | Open |
Two quirks are worth knowing before you try it. First, K3.1 Pro only works in thinking mode, so requests with thinking turned off are rejected. Second, Moonshot says it removed cyber content from its general mid-training data and from reinforcement learning environments after tracing rising exploit skill to that material, a safety edit it made before opening the weights.
The license is open, with two catches for big companies
The weights ship under Moonshot’s own Kimi K3.1 Pro License, which follows the pattern of the K3 license. For almost everyone, it is permissive: download, modify, deploy, and sell products built on the model. Two conditions bite only at large scale. A business running the model as a hosted service with more than $20 million in revenue over 12 months needs a separate agreement with Moonshot, and products above 100 million monthly users or $20 million in monthly revenue must display the model’s name. If you are a solo developer or a small team, neither condition touches you.
What this means for the open model race
The timing is the story. Just days earlier, Mistral made its 1 trillion parameter Large 4 free to download, and DeepSeek open sourced an agent that runs its own coding team. Chinese and European labs are now releasing frontier class models openly at a pace that closed labs cannot ignore, and each release raises the floor for what a free model can do. Kimi K3.1 Pro is the biggest and, on coding agent tests, the strongest of them so far.
The honest caveat is that benchmarks from the lab that built the model are a starting point, not a verdict. The weights being public is what turns a claim into something checkable. Expect independent evaluations in the coming weeks, and expect them to focus on the gap between 65.7 and 66.4 percent, because that gap is now small enough to argue about. For now, the facts are these: the weights are out, they are free for nearly everyone, and the open model race just got a new leader to chase. See also our coverage of Mistral Large 4, the 1 trillion parameter model you can run yourself, DeepSeek’s open source coding agent, and why the best AI agent scores only 12 percent on a new benchmark.
