Moonshot AI / Kimi 研究博客Source publication:
Loading content…
Moonshot AI released Kimi K3.1 Pro, the first model in its K3.1 series and its most capable 3T-class open-weight model to date, with a 1-million-token context window, alongside a Highspeed variant delivering 2-3x the speed. It is a post-trained version of Kimi K3 built with substantially larger post-training compute, supported by optimizations across the RL framework, long-context inference, sandbox infrastructure and rollout scheduling. On Terminal-Bench 4.0, which tests coding agents on realistic terminal tasks, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%.
An open-weight model now sits in the same band as the strongest proprietary models on realistic terminal tasks for coding agents. Such terminal tasks had been led by proprietary models, and K3.1 Pro moves the open-weight side close to them. On Terminal-Bench 4.0, K3.1 Pro scores 65.7%, GPT-6 Astra 58.2%, and Claude Opus 5.5 66.4%.
Long-horizon coding is strengthened through large-scale reinforcement learning, letting the model diagnose and correct problems across multi-step work. On the internal Kimi Code Bench 3.0 Long-horizon, K3.1 Pro scores 61.2%, above K3's 55.95%. In one internal test it built a walking-dynamics simulator on human-atlas, an anatomical model with 296 skeletal and 402 muscle components, and across 167 missions and 420 independent reviews iterated the controller until the simulated human went from collapsing within seconds to walking with alternating strides for 11.2 seconds.
The model extends to 3D modeling and financial workflows, with targeted work on factual accuracy. On BenchCAD, which tests CAD code generation from multi-view renderings, K3.1 Pro scores 92.7% with tools, approaching Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%; on OfficeQA Pro v2 it reaches a state-of-the-art 64.4. On the internal Kimi Hallu Bench Hard stress test, its claim-level hallucination rate fell from 23.4% for K3 to 8.3%, and the share of responses with at least one factual error was 42.4% of its predecessor's rate.
The report covers the benchmark scores and internal tests published by the releasing organization, including Terminal-Bench 4.0, Kimi Code Bench 3.0 Long-horizon, BenchCAD, OfficeQA Pro v2 and Kimi Hallu Bench Hard; full model weights are scheduled for release by Nov 4, and enterprise administrators have access off by default at launch.
The releasing organization notes that Kimi Hallu Bench Hard deliberately probes error-prone situations, so the reported rates do not represent overall error rates in daily use; its offensive cyber capability remains behind frontier proprietary models that have been publicly available for at least three months.
