Moonshot AI
Moonshot AI publishes the full weights of Kimi K3.1 Pro
The 2.8-trillion-parameter model that Moonshot’s own tests place close to the best closed systems is now downloadable, under a licence with two conditions for very large operators.
On 4 November 2026 Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face, keeping the date it set when the model launched in its apps and API in October. Anyone can now download a model that Moonshot’s benchmarks place within a few points of the strongest closed systems on several agent tests.
What was released
Kimi K3.1 Pro is a mixture-of-experts model with 2.8 trillion parameters in total, of which 104 billion are active for any given token. It reads text and images, has a context window of about one million tokens, and was trained with quantisation built in, so the weights ship in MXFP4. The repository holds 96 weight files totalling roughly 1.56 TB: this is a data-centre model, not a laptop one. Moonshot recommends the vLLM, SGLang or TokenSpeed inference engines to serve it, and it is also available through the Kimi API and Hugging Face inference providers.
Kimi K3.1 Pro is an open-weight, natively multimodal agentic model, and our most capable 3T-class model to date.
How it compares, by Moonshot’s numbers
- Terminal-Bench 4.0: 65.7, against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra.
- Agents’ Last Exam (all tasks): 43.2, ahead of Opus 5.5 at 42.6 and GPT-6 Astra at 41.8.
- GDPval-AA, an Elo rating for real-world knowledge work: 1846, behind Opus 5.5 at 1866.
- OSWorld 2.1, a computer-use test: 79.6, behind Opus 5.5 at 81.8 and Claude Fable 5.1 at 80.7.
The same table shows where it trails. On ExploitBench, a security test, it scores 74.7 against 100 for GPT-6 Astra; Moonshot says it deliberately removed cyber-related data from mid-training and reinforcement learning. It also reports that the claim-level hallucination rate on an internal test fell from 23.4% for Kimi K3 to 8.3%. All of these are the vendor’s results, and independent evaluations will take time.
Two licence conditions to read
The Kimi K3.1-Pro License is permissive for most users, with two conditions. A company that runs a model-as-a-service business and earns more than US$20 million over any 12 months must sign a separate agreement with Moonshot before commercial use. A product with more than 100 million monthly users, or more than US$20 million in monthly revenue, must display ‘Kimi K3.1-Pro’ in its interface. Purely internal use is exempt from both.
What it means for your business
Few businesses will host a 1.56 TB model themselves. The practical change is choice: a model near the top of the agent benchmarks can now be served by any inference provider with the hardware, in any region, under terms you can read in advance. That puts pressure on prices and gives you a fallback if one vendor changes its terms.
For teams that run work through AI agents, the lesson is to keep workflows independent of any single model. If your processes, approvals and records live in your own systems, swapping the model underneath is a configuration change, not a rebuild. That’s how you stay on top of the business rather than inside it: your people decide what gets done and who signs off, and the model is a part you can replace.
Sources
- Kimi K3.1 Pro model card (opens in a new tab) · Moonshot AI on Hugging Face
- Kimi K3.1 Pro (opens in a new tab) · Moonshot AI (Kimi blog)
- Kimi K3.1-Pro License (opens in a new tab) · Moonshot AI on Hugging Face
- Moonshot AI releases Kimi K3.1 Pro weights for its 2.8-trillion-parameter model (opens in a new tab) · Runtime Wire
General information only, not legal, financial or professional advice. Check the sources for your circumstances.