From the source
Moonshot AI introduced Kim K3.1 Pro, a post-trained upgrade of the Kimi K3 algorithm it released in July, which the company says can outperform Claude Opus 5.5 and GPT-6 Astra on some tasks.
- Kimi K3 comprises 896 neural networks with 3 trillion parameters and activates only 16 at a time to minimize inference costs, and the new model presumably carries over its Kimi Delta Attention memory-saving feature.
- Moonshot AI says the post-training workflow used significantly more processing power than past projects and ran on an upgraded post-training stack with automated quality checks.
- Kim K3.1 Pro scored 92.7% on the BenchCAD 3D model coding benchmark, within four percentage points of Astra and Opus 5.5, and 65.7% on Terminal-Bench 4.0, more than 7% ahead of Astra but slightly behind Opus 5.5.
- It also outperformed both on OfficeQA Pro v2, a test of complex financial analysis, while trailing them on cybersecurity benchmarks because Moonshot AI says it deliberately removed cybersecurity-related information from the post-training dataset to address safety risks.
- The model is available through Moonshot AI's paid cloud services on launch, and the company plans to open-source its weights by Nov. 4.