Moonshot Releases Kimi K3.1 Pro Weights Openly
The 2.8-trillion-parameter open-weight model is now downloadable on Hugging Face, the largest open-weight release by total parameter count.
Updated on Nov. 5, 2026 in Artificial Intelligence
Live Poll
Do you think free access to powerful AI models will benefit you?
Moonshot AI published the full weights for its Kimi K3.1 Pro model on Hugging Face, making the 2.8-trillion-parameter system available without a waitlist or gate. The release follows the model's first API launch on October 23.
Why it matters
Open weights let developers and companies run a frontier-scale model on their own hardware rather than relying solely on paid API access. The release adds pressure on proprietary AI providers whose models K3.1 Pro benchmarks against.
The repository holds 1.56 terabytes across 96 safetensors shards, and running the model requires about eight Nvidia B300 GPUs with 2.3 terabytes of combined GPU memory. Only about 104 billion of its 2.8 trillion parameters are active per token, a mixture-of-experts design.
The players
Moonshot AI
A Chinese artificial intelligence company that developed the Kimi model family and released K3.1 Pro.
Hugging Face
An open-source platform that hosts machine learning models and datasets, where the K3.1 Pro weights were published.
The details
Moonshot AI released K3.1 Pro under a license requiring hosted API providers with over $20 million in trailing 12-month revenue to sign a commercial agreement with the company. The model supports a 1,048,576-token context window and accepts images alongside text.
Timeline
October 23, 2025: Moonshot first offered K3.1 Pro through an API.
November 4, 2026: Moonshot published the full model weights on Hugging Face.
Tech Race
The K3.1-Pro License's $20 million revenue threshold for hosted API providers sets a commercial condition on an otherwise open release. The model's 2.8 trillion parameters also exceed DeepSeek V4 Pro's 1.6 trillion, making it the largest open-weight model by total parameter count so far.
Developers can download the K3.1 Pro weights directly and self-host the model if they have the required multi-GPU hardware. Companies offering the model as a hosted service above the $20 million revenue threshold must sign a commercial agreement with Moonshot AI first.
The takeaway
Open-weight releases at this scale let organizations evaluate a frontier-class model without per-call API fees, but the hardware demands mean self-hosting is practical mainly for well-resourced teams. Anyone planning to offer K3.1 Pro commercially should check whether the license's revenue threshold applies to them.
Further reading
Explore more reporting in the Artificial Intelligence section for broader context on open-weight models.
More information
The complete model weights are available on Kimi K3.1 Pro weights on Hugging Face.
Live Poll
Do you think free access to powerful AI models will benefit you?
