Moonshot released the Kimi K3.1 Pro weights on Hugging Face

Moonshot posted the weights of Kimi K3.1 Pro on Hugging Face three days ago. The company had said on 23 October that the weights would be released by Nov 4. A reader on Hacker News flagged the upload earlier today, and the post reached 232 points.

Moonshot describes Kimi K3.1 Pro as a mixture-of-experts model. Total parameters are 2.8 trillion. Activated parameters are 104 billion. The model has 93 layers and 896 experts. It picks 16 experts per token. It has two shared experts. The context window is 1,048,576 tokens. The weights come quantised to MXFP4, with MXFP8 activations, from quantisation-aware training.

What the architecture does

Moonshot replaced the full-attention layers used in Kimi K3 with Gated DeepSeek Sparse Attention. It says the change makes the model faster on long inputs. The layer mix is 69 KDA layers and 24 Gated DSA layers. The model reads text and images. The vision encoder is MoonViT-V2 at 401 million parameters. Moonshot states that Kimi K3.1 Pro was post-trained from Kimi K3. It used large-scale joint reinforcement learning across hundreds of domains.

Running it

The model card lists vLLM, SGLang and TokenSpeed as the suggested inference engines. It also lists Transformers and Docker Model Runner as other entry points. The full weights are split across 96 safetensors shards. Most shards are 17 gigabytes each. A self-host test needs a very large graphics memory budget.

Benchmark numbers Moonshot reports

Every figure below is Moonshot's own and comes from the model card, with no independent run yet.

On Terminal-Bench 4.0, Moonshot reports 65.7 for its model against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. On SWE Marathon v1.1, it reports 51.7 against 46.3 for Claude Opus 5.5 and 42.5 for GPT-6 Astra. On BenchCAD with tools, it reports 92.7 against 96.2 for Claude Opus 5.5 and 95.9 for GPT-6 Astra. On the internal Kimi Code Bench 3.0 long-horizon split, it reports 61.2 against 58.5 for Claude Opus 5.5.

For agent work, Moonshot reports 70.0 on HLE-Full with tools against 67.7 for Claude Opus 5.5. It reports 79.6 on OSWorld V2.1 against 81.8 for Claude Opus 5.5. On GPQA-Diamond, it reports 94.2 against 93.9 for Claude Opus 5.5 and 96.1 for GPT-6 Astra.

Moonshot also reports an internal hallucination score, where lower is better: 8.3 for Kimi K3.1 Pro against a range of 5.5 to 23.3 for the compared models.

The licence is not Apache

The repository carries a custom Kimi K3.1 Pro Licence. It grants broad rights to use, copy, modify and redistribute. It then names two commercial triggers. A company that runs a Model as a Service business needs a separate agreement with Moonshot. The trigger is 20 million US dollars in revenue across any 12 months. A commercial product with more than 100 million monthly active users must display "Kimi K3.1-Pro" in its interface. The same rule applies once a product's monthly revenue tops 20 million US dollars. Internal use and use through Moonshot's own products are exempt.

Hugging Face lists 797 downloads of the model in the last month, as of today.

Source

Kimi K3.1 Pro on Hugging Face, with the Hacker News discussion.