Kimi K3.1 Pro's open weights are out on the day Moonshot promised, and it is now the top open model on our index

Moonshot published the full weights of Kimi K3.1 Pro on Hugging Face on November 4, the date it gave when the model launched in its apps and API on October 23. The repository, moonshotai/Kimi-K3.1-Pro, is public and not gated. Hugging Face shows it was created on November 2; it was not in Moonshot's public model list when we checked around 14:30 UTC today, and it was there at 15:31 UTC.

THE FILES

96 safetensors shards, 1.56 TB in all, the same size as Kimi K3's release. Per Moonshot's model card the weights are MXFP4 with MXFP8 activations and the model was trained that way (quantization-aware training), so this is the precision Moonshot itself serves, not a later compression. 2.8 trillion total parameters, 104 billion active per token, a 1,048,576-token context, text and image input. The card recommends vLLM, SGLang or TokenSpeed to run it.

WHAT CHANGED FROM K3

Per the model card, K3.1 Pro is post-trained from Kimi K3, and the architecture changes in one place: the full-attention layers become Gated DeepSeek Sparse Attention layers, for cheaper long context (69 KDA and 24 Gated DSA layers, 93 in all).

THE LICENSE

The Kimi K3.1-Pro license is the Kimi K3 license word for word, with the name changed; we compared the two files. Use, modification, fine-tuning and redistribution are free, with two conditions. A company that sells model access as a service, an API for example, with more than $20 million in revenue over any 12 months needs a separate agreement with Moonshot before commercial use. A product with more than 100 million monthly users or more than $20 million in monthly revenue must display "Kimi K3.1-Pro" prominently. Neither applies to internal use.

WHERE IT RANKS

On the ArtificialWatch Bench Index, the median percentile across the public leaderboards that score it, Kimi K3.1 Pro reads 69.9, #11 of 44 models, open and closed. That makes it the highest-ranked model with downloadable weights, just ahead of DeepSeek V4.1 Flash at 68. On our coding index it is #21 of 45 at 45.1, behind four other open models. Its model page and the open-weights leaderboard now list the repository.

WHAT WE DO NOT KNOW

Whether Kimi K3.1 Pro Highspeed, the 2-3x faster version Moonshot launched alongside it, is a separate checkpoint. Moonshot has not said, and no Highspeed repository is on Hugging Face.