Beijing-based artificial intelligence developer Moonshot AI announced Kimi K3.1 Pro on October 23, 2026, positioning the flagship 3-trillion-parameter-class model as its most capable open-weight architecture to date. Serving as a heavily post-trained evolution of Kimi K3, the new model pairs a 1-million-token context window with multi-modal reasoning across text, high-resolution imagery, and video. Along with the primary release, Moonshot unveiled Kimi K3.1 Pro Highspeed, which delivers two to three times faster inference throughput for interactive coding and terminal workflows across Kimi.ai, Kimi Work, and the Kimi API.
The technical gains achieved by K3.1 Pro are anchored in an overhaul of Moonshot's reinforcement learning infrastructure. Rather than scaling pre-training data volume alone, the lab optimized its entire post-training stack—spanning distributed RL rollouts, long-context KV cache management, and automated sandbox validation—to coordinate joint reinforcement learning across hundreds of downstream domains. On Terminal-Bench 4.0, which challenges autonomous coding agents with complex terminal environments and rigid verification tests, K3.1 Pro posted a 65.7% success rate, surpassing OpenAI's GPT-6 Astra at 58.2% and closing within a single percentage point of Anthropic's Claude Opus 5.5 at 66.4%.
Beyond terminal programming, K3.1 Pro demonstrated significant capabilities in spatial geometry and parametric engineering. On BenchCAD, an evaluation measuring 3D reconstruction from multi-view sketches into executable CAD scripts, K3.1 Pro scored 92.7% when equipped with tool-use connectors, approaching GPT-6 Astra's 95.9%. In practical demonstrations, the model generated interactive Three.js scenes, procedural Godot game mechanics, and designed mechanical components directly in Blender without requiring external MCP bridge middleware. On financial analytics, the model posted 64.4 on OfficeQA Pro v2, cutting factual calculation errors by more than half relative to baseline K3.
Addressing safety governance, Moonshot AI detailed an explicit defensive policy governing offensive cyber capabilities. During development, researchers observed that scaling post-training compute naturally elevated exploit discovery proficiency on software like the Linux kernel and JavaScriptCore. In response, Moonshot systematically identified and purged offensive cybersecurity data from both mid-training and RL environments. On xdev-bench-hard, an unmitigated research checkpoint solved 90.5% of exploit scenarios at pass@3, whereas the shipping K3.1 Pro was constrained to 81.0%, keeping its offensive threshold strictly below frontier proprietary systems while leaving open weights accessible for defensive auditing on November 4.