Moonshot Releases Kimi K3.1 Pro Weights on Hugging Face, Delivering on November 4 Promise

Moonshot AI kept its word. On November 4, 2026, the self-imposed deadline set at launch, the Chinese company released on Hugging Face the full weights of Kimi K3.1 Pro, its flagship model debuted on October 23. We highlighted this commitment in our article on its entry into the Intelligence Index: it has now materialized. K3.1 Pro thus becomes one of the largest downloadable models to date.

What Has Been Released

The model card describes a Mixture-of-Experts with 2.8 trillion parameters, including 104 billion active per token, spread across 896 experts (16 routed per token). The context window reaches 1,048,576 tokens—one million—and the model accepts text and images thanks to a 401-million-parameter vision encoder. Compared to Kimi K3, Moonshot replaced its full attention layers with DeepSeek-inspired sparse attention to better handle long contexts.

The weights arrive pre-quantized (MXFP4 for weights, MXFP8 for activations), an approach built in during training rather than applied as an afterthought. Even so, the repository totals around 1.56 TB according to the file list on Hugging Face: this is far from a model that can run on a workstation. Moonshot recommends three inference engines—vLLM, SGLang, and TokenSpeed—and continues to offer its API (kimi-k3.1-pro) for those who prefer not to host it.

An Open License With Two Guardrails

The “Kimi K3.1 Pro License” largely mirrors the text of an MIT license: use, modification, redistribution, sale, and derivative works are permitted free of charge. Two caveats target massive commercial players:

• a company that resells model access via an API (a “model service”) and exceeds $20 million in revenue over 12 months must sign a separate agreement with Moonshot;
• a commercial product that exceeds 100 million monthly active users or $20 million in monthly revenue must prominently display “Kimi K3.1-Pro”.

These stipulations apply neither to internal enterprise use nor to access via Moonshot’s official products or certified inference partners. For a company seeking to self-host the model for internal teams—much like Revolut already does with GLM—the license presents no hurdle.

Reported Performance Approaching Top Closed Models

Moonshot positions K3.1 Pro as neck-and-neck with the leading proprietary models in coding and agentic workflows. On Terminal-Bench 4.0, it claims 65.7, compared to 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra; on DeepSWE v1.1, it posts 75.5. The model card table also shows clear deficits elsewhere, such as on HLE-Diamond (39.4 versus 60.9 for Opus 5.5). These figures come from the developer, obtained using its own settings and varied harnesses across models; only some scores are drawn from third-party leaderboards.

Deliberately Curtailed Cyber Capabilities

Notably, Moonshot explains that it removed cybersecurity-related data from the training of this open-weights version. On its internal vulnerability discovery and exploitation benchmark, the released model successfully completes 81.0% of tasks, compared to 90.5% for a version trained without this filtering. The company claims it does not wish to release weights that push the frontier of offensive capabilities, while acknowledging that general coding advancements also benefit such tasks. These metrics remain internal and have not been validated by a third party.