Moonshot AI Just Put 2.8 Trillion Parameters on Hugging Face

On November 4, 2026, Moonshot AI pushed the full model weights of Kimi K3.1 Pro to Hugging Face, making one of the largest open-weight language models publicly available. The release follows the API launch from roughly two weeks prior.

The numbers are significant. Kimi K3.1 Pro uses a Mixture-of-Experts architecture with 2.8 trillion total parameters — but only 104 billion of those activate per inference token. That tradeoff is intentional: MoE models pack theoretical capacity without proportionally scaling compute costs at inference time. The model's context window sits at roughly 1,048,576 tokens (about 1 million), putting it among the longest contexts available from an open-weight model.

Downloading it requires planning. The weights ship as MXFP4-quantized tensors across 96 safetensors shards, totaling approximately 1.56 terabytes. Running Kimi K3.1 Pro locally is not a weekend laptop project — it requires a serious GPU cluster or a managed serving setup using vLLM 0.34.0+, SGLang, or TokenSpeed.

The model accepts both text and image input, putting it in the multimodal open-weight tier alongside other recent heavy releases. For those who can't run it locally, the API is available at platform.kimi.ai as kimi-k3.1-pro.

The license deserves careful reading before you build anything on it. Moonshot ships the weights under its own custom Kimi K3.1 Pro License — described by multiple observers as MIT-style, but with meaningful additions. Large hosted API services and very large product deployments face additional commercial conditions. For research use or smaller-scale deployment, the license is likely permissive enough, but commercial operators should check the license file directly before committing to it.

The release cements Moonshot, a Beijing-based company, as one of the more credible participants in the race to put frontier-scale models into the open-weight ecosystem. After the API-only period, the full weights landing on Hugging Face marks a meaningful step toward letting the broader research community benchmark, fine-tune, and scrutinize the model independently — which is ultimately how trust in these systems gets established.

The question now is who can actually run it. At 1.56 TB, Kimi K3.1 Pro sits in the tier of models that only well-resourced teams or cloud instances can serve. But in a landscape where most frontier performance comes from closed APIs, the fact that the weights are public matters regardless of the hardware bar.