Moonshot AI launches Kimi K3.1 Pro, a 3-trillion-parameter flagship model

Kimi K3 comprises 896 neural networks with 3 trillion parameters but activates only 16 of them at any given time to reduce inference costs. Moonshot AI says the post-training run for the new version used significantly more processing power than its earlier projects, and it ran on an upgraded post-training stack with automated quality checks and better infrastructure efficiency.

The company highlights 3D modeling, where the model writes code that defines geometry and can also animate the resulting models. It scored 92.7% on the BenchCAD benchmark, within four percentage points of GPT-6 Astra and Claude Opus 5.5, and reached 65.7% on Terminal-Bench 4.0, more than 7% ahead of Astra but slightly behind Opus 5.5. It also outperformed both rivals on OfficeQA Pro v2, a test of complex financial analysis.

Performance on popular cybersecurity benchmarks trails Astra and Opus 5.5 by a significant margin. According to Moonshot AI, its engineers removed cybersecurity-related information from the post-training dataset to address safety risks, and the company plans to add more advanced vulnerability analysis features to future models.

Developers who need a model for code-driven 3D generation or financial analysis now have another option to test, and open weights are promised for later this month. The figures are Moonshot AI's own benchmark results, so teams should run their own evaluations before switching. Agents that depend on security analysis should not assume Kimi K3.1 Pro matches the top closed models in that area.