Moonshot AI releases Kimi K3.1 Pro open-weight model with 2.8T parameters

• Moonshot AI released Kimi K3.1 Pro, an open-weight native multimodal agentic model with 2.8T total parameters and 104B active parameters.
• The model supports a 1,048,576 token context window and uses Gated DeepSeek Sparse Attention (Gated DSA) to improve long-context efficiency.
• Benchmark results show 65.7 on Terminal-Bench 4.0, 51.7 on SWE Marathon v1.1, and 94.2 on GPQA-Diamond.
• Weights are available on Hugging Face under the Kimi K3.1 Pro License, with API access via platform.kimi.ai and inference support for vLLM, SGLang, and TokenSpeed.

Architecture and Specifications

• Mixture-of-Experts (MoE) architecture with 93 layers, 896 experts, and 16 selected experts per token.
• Attention configuration includes 69 KDA and 24 Gated DSA heads with a hidden dimension of 7168.
• Vision encoder uses MoonViT-V2 with 401M parameters; quantization supports MXFP4 weights and MXFP8 activations.

Evaluation and Deployment

• Coding benchmarks were evaluated using the Claude Code harness, with BenchCAD sampled from 1,000 of 17,900 tasks.
• Agentic benchmarks include HLE-Full (70.0), BrowseComp (92.1), and Automation Bench-AA (73.3).
• API usage requires passing previous assistant reasoning_content and tool_calls in multi-turn conversations, with reasoning_effort set to max by default.