tl;dr: Moonshot released open weights for Kimi K3.1 Pro, a 2.8T-parameter MoE model (104B activated) with a 1M-token context, natively multimodal input, and Gated DeepSeek Sparse Attention replacing the prior Gated MLA layers for long-context efficiency. Benchmarks show it competitive with Claude Opus 5.5, GPT-6 Astra, and Claude Fable 5.1 on coding and agentic tasks (e.g., leading on SWE Marathon, BrowseComp, and HLE-Full with tools), while trailing on some reasoning benchmarks like HLE-Diamond. It ships with MXFP4 quantization-aware training and is deployable via vLLM, SGLang, and TokenSpeed.