Kimi K3.1 Pro model weights are out
Moonshot AI has published the full weights for Kimi K3.1 Pro on Hugging Face, in a public, ungated repository named moonshotai/Kimi-K3.1-Pro. The model itself is not new. Moonshot launched it on October 23 as an API and app model, and its launch announcement had set November 4 as the target for releasing the weights. The repository holds about 1.56 TB spread across 96 safetensors shards. The weights use MXFP4 quantization with MXFP8 activations, and training used quantization-aware methods.
Kimi K3.1 Pro is a mixture-of-experts model with 2.8 trillion total parameters, of which about 104 billion activate per token. Its context window is about 1 million tokens, and its model card lists text and image inputs. The hosted API continues to run, with the model ID kimi-k3.1-pro and OpenAI- and Anthropic-compatible interfaces. Launch pricing was $2 per million input tokens and $10 per million output tokens, a third below Kimi K3.
According to BigGo Finance, Moonshot removed attack-related content from its mid-training and reinforcement learning tasks, which lowered the pass rate on vulnerability exploitation tasks from 90.5% to 81.0%. On the Agent Arena leaderboard, the model ranks first among open-weight models and 13th overall, with a net improvement of +5.1%. Artificial Analysis gives it a score of 46 on its Intelligence Index v5.0.1.
Why it matters
For teams building on AI models, open weights change the cost question. A 2.8 trillion parameter model with a million-token context can now be inspected, fine-tuned and self-hosted, which removes the per-token API bill for teams able to run it. The hardware bar is high, though. At roughly 1.56 TB, the weights require a multi-GPU server with large amounts of memory, so infrastructure budgets need to be set before any plans are made around self-hosting.
For investors and sellers, a near-frontier model that anyone can download puts pressure on the pricing of closed model vendors and on the margins of API resellers. Benchmark results remain company-reported and sources disagree on some of them, so buyers should run their own evaluations. License terms also need review before commercial use, since the license is custom rather than a standard open source license and reports describe its conditions differently.
Key points
- 1Moonshot published Kimi K3.1 Pro weights on Hugging Face on November 4, 2026.
- 2The model has 2.8 trillion total parameters, with about 104 billion active per token.
- 3The weights total roughly 1.56 TB across 96 safetensors shards.
- 4The model launched through Moonshot's API on October 23, before the weights were released.
- 5Moonshot's removal of attack-related training data cut a vulnerability exploitation pass rate from 90.5% to 81.0%.
