Open-Weight AI Model Kimi K3.1 Pro Released by Moonshot AI
Chinese AI developer Moonshot AI has published Kimi K3.1 Pro, an open-weight multimodal model with 2.8 trillion parameters and a one-million-token context window. Built on Kimi K3, it is said to improve coding and agentic performance.
Moonshot AI, the developer behind the Kimi family of language models, has released the weights of Kimi K3.1 Pro on Hugging Face. The model is open-weight, meaning its trained parameters can be downloaded publicly. It is natively multimodal, handling both text and images, and is designed for agentic tasks in which an AI system carries out multi-step work on its own. The developers call it their most capable model in the roughly three-trillion-parameter class.
Architecture
Kimi K3.1 Pro uses a Mixture-of-Experts design: it has 2.8 trillion parameters in total, with 104 billion activated for each token. It contains 896 experts, 16 of which are selected per token, plus two shared experts. The context window is 1,048,576 tokens. Compared with Kimi K3, the full-attention layers were replaced with Gated DeepSeek Sparse Attention layers to handle long inputs more efficiently. Image processing relies on the 401-million-parameter MoonViT-V2 encoder, and the weights are quantized in MXFP4 format.
Results
The model is the first in the K3.1 series and was produced by further training Kimi K3 with large-scale reinforcement learning across hundreds of domains. The developers cite improvements in coding, 3D creation, financial research and factual accuracy. Their published benchmark tables compare it with Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017 and GLM-5.3. On Terminal-Bench 4.0, for example, Kimi K3.1 Pro scores 65.7, while Claude Opus 5.5 scores 66.4. On other tests, including Program Bench, it trails competitors. The results were largely produced by the developers themselves, and some benchmarks are internal.
Availability
The model is available through the API at platform.kimi.ai by selecting kimi-k3.1-pro; the API is compatible with OpenAI and Anthropic formats. Recommended inference engines are vLLM, SGLang and TokenSpeed. Thinking effort is set with the reasoning_effort field, which accepts "low", "high" and "max", with "max" as the default.
