Kimi K3.1 Pro

Kimi K3.1 Pro is Moonshot AI's most capable 3T-class open-weight model, post-trained from Kimi K3 with substantially larger reinforcement-learning compute. It keeps the K3 architecture — 2.8 trillion total parameters across 93 layers, 16 of 896 experts active per token for roughly 104 billion active parameters, Kimi Delta Attention paired with gated DSA, and a MoonViT-V2 vision encoder — with MXFP4 weights and MXFP8 activations from quantization-aware training. Against K3 it lifts Kimi Code Bench 3.0 Long-horizon from 55.95 to 61.2 and cuts claim-level hallucination on Kimi Hallu Bench Hard from 23.4% to 8.3%, and it scores 65.7 on Terminal-Bench 4.0, 92.1 on BrowseComp, 94.2 on GPQA-Diamond and 79.6 on OSWorld V2.1. Thinking is always on, with selectable low, high, or max reasoning effort, making it suited to long-horizon agent runs, repository-scale engineering, financial research, and document-heavy multimodal work.