Moonshot AI open-weights Kimi K3.1 Pro
On 4 Nov 2026 Moonshot AI (Kimi) said it is releasing the model weights for flagship Kimi K3.1 Pro on the official Hugging Face account moonshotai. The open-weight, natively multimodal agentic model is a ~2.8T-parameter MoE (≈104B activated) with a 1M-token context; API access remains available as kimi-k3.1-pro on platform.kimi.ai.
Verified against the official @Kimi_Moonshot post (2026-11-04T16:00:35Z) and the official model card moonshotai/Kimi-K3.1-Pro: Moonshot AI announced it is releasing the model weights of Kimi K3.1 Pro, adding that it will keep advancing open intelligence.
What it is
Per the card, Kimi K3.1 Pro is an open-weight, natively multimodal agentic model and Moonshot’s most capable ~3T-class model to date. It is post-trained from Kimi K3 with large-scale joint reinforcement learning across many downstream domains, and is the first model in the K3.1 series. The card says it improves coding, 3D creation, financial research, factual accuracy, and other agentic capabilities, with a 1-million-token context window.
Scale and architecture (as stated on the card)
- Architecture: Mixture-of-Experts (MoE)
- 2.8T total parameters; 104B activated
- 896 experts, 16 selected per token, plus 2 shared experts
- Context length 1,048,576 tokens
- Modalities: text and image; vision encoder MoonViT-V2 (~401M parameters)
- Quantization: MXFP4 weights / MXFP8 activations (quantization-aware training)
How to use it
- Weights: Hugging Face repo
moonshotai/Kimi-K3.1-Pro - API: select
kimi-k3.1-proon platform.kimi.ai; OpenAI/Anthropic-compatible APIs are provided - Recommended inference stacks on the card: vLLM, SGLang, TokenSpeed
- Thinking effort: top-level
reasoning_effortsupportslow/high/max(defaultmax)
License
Released under the Kimi K3.1 Pro License named on the model card.
This summary sticks to the official card and post; vendor benchmark claims are not treated as independent third-party verification.