Kimi K3.1 Pro's Open Weights: A 2.8-Trillion-Parameter Model You Can Download, If You Can Store 1.56 TB
Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face on November 4, keeping a date it had promised when it launched the model in its apps and API in late October. The repository is public, with no application gate.
What you get
- Size: 2.8 trillion parameters in total, with 104 billion active at any time (a mixture-of-experts design).
- Context: a 1-million-token context window, with text and image input.
- Format: MXFP4 weights. Analysts who inspected the repository put the files at about 1.56 TB.
- Serving: the model card points to inference engines such as vLLM and SGLang.
The licence: open, but with conditions
The weights are released under the Kimi K3.1 Pro License. It grants permission to use, modify and distribute the software, subject to its conditions. Model-as-a-service businesses whose aggregate revenue exceeds $20 million over any consecutive 12 months must enter a separate agreement before commercial use. Commercial products above specified user or revenue thresholds must display the model name. The licence states that these two requirements do not apply to defined internal use or use through Moonshot's official products or certified inference partners. Prospective users should read the full licence and obtain advice appropriate to their use.
Cyber data and the limits of Moonshot's test
In its technical blog, Moonshot says it removed cyber-related material from mid-training data and filtered cyber tasks from reinforcement-learning environments. Its internal xdev-bench-hard test reports a pass@3 result of 90.5% for a checkpoint without these measures and 81.0% for the final model. The benchmark is internal; these are the company's own results. They do not establish whether fine-tuning could restore capability.
Our analysis
Open weights allow inspection and local deployment, subject to the model licence and the hardware available. The roughly 1.56 TB download is substantial, and deployment also requires working memory and compute. Teams planning commercial use should read the licence directly. The release came one day before OpenAI announced restricted access to GPT-6 Cyber; the two releases take different approaches to access and cyber capability. The effectiveness of either approach is a separate question.
Sources: