Europa Press reported that Moonshot AI’s Kimi K3.1 Pro weights became available on Hugging Face on November 4, 2026. Moonshot had announced the model on October 23 and set November 4 as its release deadline. The downloadable weights open a path to running the model on infrastructure chosen by its users; Moonshot’s model card names serving engines and spells out an important requirement for multi-turn API calls.
For a broader look at Chinese open-weight models, see NeoTeo’s earlier coverage. Kimi K3.1 Pro is a separate release.
Treat downloading the weights and meeting the license terms as two separate deployment decisions.
What the model card says about local serving
Kimi K3.1 Pro uses a mixture-of-experts architecture, a design that activates selected groups of model parameters for each token. Moonshot’s model card lists 2.8 trillion total parameters, 104 billion activated parameters, 896 experts, 16 selected experts per token, and two shared experts. It also specifies a context length of 1,048,576 tokens. The repository hosts the model card and files for the release.
Moonshot recommends vLLM, SGLang, and TokenSpeed as inference engines and includes serving examples in its Hugging Face model card. Those names give developers a starting point for choosing software to serve the weights; they don’t determine a system’s hardware configuration.
The card also gives a detail that matters when building an API integration. For multi-turn conversations and tool calls, the complete assistant message returned by the API must be passed back into the next request, including reasoning_content and tool_calls, rather than just the visible content. The reasoning_effort setting accepts low, high, or max; its default is max.
How to read the Terminal-Bench comparison
Moonshot reports a 65.7% score for Kimi K3.1 Pro on Terminal-Bench 4.0. Its published comparison lists 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra.
| Model | Terminal-Bench 4.0 score reported by Moonshot AI |
| Kimi K3.1 Pro | 65.7% |
| Claude Opus 5.5 | 66.4% |
| GPT-6 Astra | 58.2% |
Moonshot says it evaluated Kimi K3.1 Pro with the Claude Code harness. The other scores are the maximum values reported in model blog posts or on the Terminal-Bench leaderboard, so the table brings together results with mixed provenance rather than one controlled, independent head-to-head. Moonshot’s model card says its results generally use maximum reasoning effort, temperature 1.0, and top-p 0.95 unless a benchmark’s notes specify otherwise. The company’s benchmark announcement gives the scores and evaluation context.
A separate English-language video by VitoMind walks through builds from eight matched prompts. The creator shows one build per prompt, without retries or agents; the models used different reasoning settings and were tested on different dates.
The weights have a model-specific license
Hugging Face names the license Kimi K3.1 Pro License. Europa Press reports that it permits use, modification, fine-tuning, and redistribution subject to conditions. Its summary says an organization offering the model to third parties as a service and billing more than $20 million annually must sign a separate agreement with Moonshot. It also reports an interface attribution requirement for products with more than 100 million monthly users or more than $20 million in monthly revenue.
Those conditions make the license name worth checking before building a service around the model. Downloadable weights provide a deployment option; the license still defines terms for use and redistribution.
