Kimi AI has released Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed, expanding its K3.1 model series. According to the announcement, K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute and a 1-million-token context window. K3.1 Pro Highspeed delivers 2-3× the speed of K3.1 Pro for faster interactions. Both models are available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, with full model weights to be released by November 4.
The company optimized multiple layers of its infrastructure—the reinforcement learning framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and environment quality checks—to enable large-scale joint reinforcement learning across hundreds of downstream domains.
On coding benchmarks, K3.1 Pro scored 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra (58.2%) but slightly behind Claude Opus 5.5 (66.4%). The model excels in long-horizon coding tasks including game engine work across Unity, Unreal Engine, and Godot. According to the post, K3.1 Pro built a human walking dynamics simulator, iteratively refining a walking controller through error diagnosis and reinforcement learning until the simulated human achieved 11.2 seconds of continuous walking.
For 3D design, K3.1 Pro scored 92.7% on BenchCAD, which tests 3D reconstruction from multi-view renderings, approaching Claude Opus 5.5’s 96.2% and matching GPT-6 Astra’s 95.9%. The model demonstrates strength in parametric modeling and can write and run scripts for Blender and CAD software.
On finance tasks, K3.1 Pro achieved 64.4 on OfficeQA Pro v2, which evaluates complex fiscal data analysis. The company reports improvements over K3 on internal finance benchmarks: Financial Modeling improved 7.2 points, Financial Research improved 11.6 points, and Business Operation achieved roughly 4.2× K3’s performance. For factual accuracy, K3.1 Pro’s claim-level hallucination rate fell from 23.4% on K3 to 8.3% on Kimi Hallu Bench Hard, an internal stress test.
Regarding cybersecurity, Kimi identified and removed cyber-related data from both mid-training and reinforcement learning stages while preserving general-purpose coding capabilities, according to the announcement.
Key facts
- K3.1 Pro features a 1-million-token context window and is available on multiple Kimi platforms starting today
- K3.1 Pro Highspeed offers 2-3× faster performance than K3.1 Pro
- On Terminal-Bench 4.0, K3.1 Pro scored 65.7%, outperforming GPT-6 Astra (58.2%)
- On BenchCAD 3D reconstruction tests, K3.1 Pro achieved 92.7%
- Hallucination rate improved from 23.4% (K3) to 8.3% (K3.1 Pro) on internal stress tests
- Full model weights will be released by November 4