Kimi K3.1 Pro model weights are out

Kimi K3.1 Pro is Moonshot AI’s open-weight, text-and-image model, presented as its most capable model in the 3-trillion-parameter class. It uses a mixture-of-experts architecture with 2.8 trillion total parameters and 104 billion activated per token, and supports a 1-million-token context window. The model was post-trained from Kimi K3 using reinforcement learning across hundreds of domains. Its design combines KDA and gated sparse attention, replacing the earlier model’s full-attention layers to make long-context processing more efficient. Weights use MXFP4 quantization, with MXFP8 activations. The model card highlights coding, 3D creation, financial research, factual accuracy, and agentic work as areas of improvement.

Published benchmark results show strong but uneven performance against several commercial and open models. Kimi leads the listed comparison on measures including BrowseComp, HLE-Full with tools, and Terminal-Bench 4.0, while scoring below rivals on other tests, such as Terminal-Bench-Science, Program Bench, several reasoning benchmarks, and cybersecurity evaluations. Results vary by task and tool use, so they indicate broad capability rather than uniform superiority. The weights can be loaded through Transformers or served with vLLM, SGLang, or Docker; the examples include an OpenAI-compatible interface for image-and-text prompts. The repository lists a Kimi K3.1 Pro license, and its model card provides detailed architecture specifications and benchmark tables.