Moonshot AI has published the full weights of Kimi K3.1 Pro on Hugging Face, keeping the promise it made when the model launched on October 23. The release turns one of the largest open-weight models ever shared into something any team can download, fine-tune, and run for itself, rather than rent through Moonshot’s API. The files went public on November 4, twelve days after launch.
What shipped
The checkpoint is a mixture-of-experts model with 2.8 trillion total parameters, of which 104 billion activate for each token, spread across 896 experts with 16 selected per token. It keeps a 1-million-token context window and a 401-million-parameter vision encoder for image input. The download is 96 files totalling about 1.56 terabytes, stored as MXFP4 weights with MXFP8 activations.
The weights ship under Moonshot’s Kimi K3.1 Pro License, which reads like an MIT license with two extra conditions for the largest users. A business earning more than $20 million a year from selling model access needs a separate agreement with Moonshot, and a product past 100 million monthly users must show the Kimi K3.1 Pro name in its interface.
Internal use and use through Moonshot’s own products or certified inference partners are exempt from both conditions, so the conditions bite only at the top of the market. For nearly every team, the practical terms are simple: download the files, fine-tune them, and deploy or sell products built on them without asking Moonshot first.
The safety edit
Moonshot also edited the open weights for safety before publishing them. It says it traced rising exploit skill to cyber content in its general training data and reinforcement-learning environments, and removed both from the released checkpoint. On Moonshot’s internal exploit benchmark, the released model scores 81.0, against 90.5 for the unfiltered checkpoint that was never released.
The public cyber scores tell the same story: 74.7 on ExploitBench and 20.7 on ExploitGym, below the hosted model’s rivals on both. Moonshot’s benchmark table still claims 65.7 on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4. All of these are vendor-reported scores; no independent lab has yet rerun the tests on the downloadable files.
What it costs to run
Running the full model yourself is still a data-center job, not a laptop one. Moonshot recommends vLLM, SGLang, or TokenSpeed as inference engines, and independent estimates put the hardware bar at eight high-end GPUs. For most teams, the hosted route stays practical: Kimi’s API charges $2 per million input tokens and $10 per million output tokens.
Our read
Our read: this release matters less for who can run it today than for what it proves about the open-weight race. At 2.8 trillion parameters, Kimi K3.1 Pro is among the largest models ever published as downloadable weights, ahead of Qwen’s 2.4-trillion model and DeepSeek’s 1.6-trillion flagship. The open-closed gap is now a deployment problem, not a capability mystery.
The timing sharpens the point. Tencent shipped its Hy4 model with vision and computer use the same day, and the open-weight field is crowded with frontier-class releases arriving weeks apart. Moonshot’s bet is that giving away the weights wins developers who would otherwise build on a closed API, and the license’s light touch is designed to make that choice easy.
Why It Matters
For developers, the release removes the biggest objection to building on Kimi: vendor lock-in. A team can prototype on Moonshot’s API and, if the economics change, take the same model in-house without rewriting its prompts or tooling. For the wider field, it raises the open floor, forcing closed labs to justify their prices against a downloadable rival.
The caveat is the hardware bill. Until smaller, distilled versions appear or cheaper inference spreads, only well-funded teams will run the full model themselves, and most users will still meet Kimi K3.1 Pro through the API. The weights are open; the compute to use them is not yet cheap, and that gap decides who actually benefits.
Background
Background: Kimi K3.1 Pro launched on October 23 as a post-trained version of Kimi K3, with substantially more post-training compute behind it. Moonshot promised the full weights by November 4, and the repository’s first commit is dated November 2, with the files going public two days later. The launch itself is covered in our earlier report.
0 Comments