Kimi-K3.1-Pro: 2.8T MoE architecture, 1M token context

Kimi K3.1 Pro, released by Moonshot AI, is an open-weight multimodal model that processes text and images simultaneously. It adopts a Mixture-of-Experts structure with 2.8T total parameters, activating only 104B active parameters per token to enhance efficiency. It supports a long context window of up to 1M tokens, enabling the analysis of massive documents or codebases in a single pass.

It improves performance across agent tasks, including coding, 3D generation, financial research, and factual accuracy. In particular, it applies Gated DeepSeek Sparse Attention to improve long-context modeling efficiency compared to the previous K3. Quantization-aware training with MXFP4 weights and MXFP8 activations also optimizes memory usage.

It scored 65.7 on Terminal-Bench 4.0, demonstrating coding capabilities comparable to Claude Opus 5.5. It proved top-tier performance with a score of 70.0 on the HLE-Full w/ tools benchmark. It is compatible with major inference frameworks such as vLLM, SGLang, and Transformers, allowing immediate deployment via an OpenAI-compatible API.

It is suitable for building agent systems that require complex tool calling and multi-step reasoning. It offers high utility in development environments that demand precise code generation and data analysis while maintaining long context. It also scored 61.2 on its own benchmark, Kimi Code Bench 3.0 Long-horizon, confirming its ability to handle long-term tasks.