BigGo Finance and Startup Fortune report that Moonshot AI published the full Kimi K3.1 Pro weights on Hugging Face on November 4, 2026. Moonshot’s October 23 launch announcement had set November 4 as the planned release date. The model card spells out the model’s attention layers and lists serving instructions for vLLM and SGLang.
This follow-up builds on NeoTeo’s earlier coverage of Kimi K3.1 Pro’s cyber-risk controls and the 81.0% figure. The new model-card details focus on architecture and deployment. (BigGo Finance; Startup Fortune; Moonshot AI’s launch announcement; Moonshot AI’s model card)
How Kimi K3.1 Pro’s attention layers differ from Kimi K3
Moonshot’s model card lists 69 KDA attention layers and 24 Gated DSA attention layers for Kimi K3.1 Pro. It says Gated DSA replaces the full-attention Gated MLA layers used in Kimi K3.
The card describes Kimi K3.1 Pro as a mixture-of-experts model, a design that routes each token to selected expert components. It lists 2.8 trillion total parameters, 104 billion activated parameters, and 896 experts, with 16 selected per token and two shared experts. Its context length is 1,048,576 tokens. The card also lists text and image as supported modalities and names MoonViT-V2 as the vision encoder.
For deployment details, start with Moonshot’s serving instructions for vLLM and SGLang.
Serving instructions in the repository
Moonshot’s model card includes serving instructions for both vLLM and SGLang. Those are the repository’s named deployment options for running the weights.
