Kimi Releases K3.1 Pro Open-Weight Model Matching Proprietary AI
Kimi launched K3.1 Pro, an open-weight model with a one-million-token context window, alongside K3.1 Pro Highspeed for 2-3x faster inference. Both are immediately available on Kimi’s platforms and API, with full model weights releasing by November 4th. The model underwent extensive post-training using large-scale reinforcement learning, optimizing every infrastructure layer from the RL framework to training efficiency.
On coding benchmarks, K3.1 Pro achieves competitive results with proprietary leaders—scoring 65.7% on Terminal-Bench 4.0 (outpacing GPT-6 Astra but trailing Claude Opus 5.5 slightly). The model excels at complex, multi-step coding tasks; one internal test had it build a human movement simulator with hundreds of skeletal and muscle components, then systematically diagnose and fix dozens of errors across hundreds of test iterations. In 3D design, it reaches 92.7% on BenchCAD. Financial analysis capabilities show particular strength, with substantial benchmark improvements and hallucination rates cut from 23.4% to 8.3% on stress-test benchmarks.
Kimi applied targeted controls to limit cybersecurity risks. The team identified cyber-related training data naturally embedded in the corpus and present in RL environments, finding it drove increases in vulnerability research capability. For the open-weight release, they removed this data from both mid-training and RL stages while preserving general coding utility, aiming to reduce misuse potential without sacrificing model usefulness.
Read the full article
Continue reading at Hacker News →This is an AI-generated summary. Read the original for the full story.