Moonshot publishes Kimi K3.1 Pro weights, opening a self-hosted route for coding agents

Developers can now download and serve Kimi K3.1 Pro themselves: Moonshot AI has published the model’s weights on Hugging Face. The release puts a 2.8-trillion-parameter system within reach of outside deployment and testing, alongside company-reported coding results that approach or surpass Claude Opus 5.5 on some tests—but fall well behind on others.

What is inside the download

Moonshot describes K3.1 Pro as the first model in its K3.1 series, developed through further training of Kimi K3. The company says it used large-scale joint reinforcement learning across hundreds of downstream domains. Its stated targets include coding, 3D creation, financial research and factual accuracy, rather than software development alone.

The model uses a mixture-of-experts architecture, with 2.8 trillion parameters in total and 104 billion activated during processing. Its published context limit is 1,048,576 tokens—the amount of material it can work with in a request. The card lists text and image inputs.

A coding contender, not an across-the-board winner

ProgramBench is another weaker result: Moonshot lists 31.5 for Kimi against 65.0 for Opus. That table uses binary task scoring, counting a task only when its ProgramBench score reaches at least 95%.

The evaluation notes are essential to reading those comparisons. Kimi ran through Claude Code on the terminal and SWE Marathon tests. That is the agent harness: the software surrounding the model as it works through tasks. For competing models, Moonshot draws scores from release posts or official leaderboards on those tests.

These are therefore not uniformly controlled, same-setup comparisons. Moonshot also says its Kimi results generally use maximum reasoning effort. The table does not isolate model quality from the surrounding setup or establish that Kimi is the stronger coding model for every job.

Serving instructions, with a named license

• Transformers examples show both a high-level image-to-text pipeline and direct model loading, with trust_remote_code enabled.
• The vLLM instructions show how to start a model server and send a request through an OpenAI-compatible API.
• SGLang instructions include a server launch, a sample request and a Docker-based deployment example.

The repository names a model-specific license, kimi-k3.1-pro. Downloadable weights and permission for every intended use are separate questions; teams considering modification, redistribution or commercial deployment should check the actual terms.