Moonshot details Kimi K3.1 Pro's benchmark gains
Moonshot's technical blog says Kimi K3.1 Pro, its 3T-class open-weight model with a 1M-token context window, scores within a point of Claude Opus 5.5 on Terminal-Bench 4.0, with weights due by November 4.
As we reported earlier, Kimi K3.1 Pro placed fourth on Artificial Analysis's Intelligence Index. Moonshot has now published a technical blog on how it built the model and how it performs.
K3.1 Pro is not a new base model. It is a post-trained version of Kimi K3, trained with substantially more post-training compute. Moonshot says it reworked its RL framework, long-context inference, sandbox infrastructure and rollout scheduling to run joint reinforcement learning across hundreds of domains.
Its headline result is 65.7% on Terminal-Bench 4.0. That puts it ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%). On BenchCAD, which tests 3D reconstruction through CAD code, it scores 92.7% with tools, compared with 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra. Moonshot also claims a state-of-the-art 64.4 on OfficeQA Pro v2, a financial document analysis benchmark.
Many of the other gains are on Moonshot's own benchmarks:
- Kimi Code Bench 3.0 Long-horizon: 61.2%, up from K3's 55.95%.
- Finance: 7.2 points higher on financial modelling and 11.6 points higher on financial research.
- Kimi Hallu Bench Hard: the claim-level hallucination rate fell from 23.4% to 8.3%.
Moonshot is also releasing K3.1 Pro Highspeed, which it says runs 2–3× faster. Both models are live on Kimi.com, Kimi Code and the Kimi API. The full weights are promised by November 4. The blog also describes cutting the model's cyber capability by removing training data, but the excerpt available to us does not include those details.