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

Kimi K3 comprises 896 individual neural networks with 3 trillion parameters and activates only 16 of them at any given time to minimize inference costs. The company also reduced its models’ hardware footprint with Kimi Delta Attention, a custom implementation of Gated DeltaNet that cuts the RAM used by the attention mechanism to interpret user prompts. The feature was presumably carried over to Kimi K3.1 Pro.

Moonshot AI said the new model’s post-training workflow consumed significantly more processing power than similar past projects. It ran on an upgraded version of the company’s post-training stack, which added automated quality checks and improved the underlying infrastructure’s efficiency.

Kimi K3.1 Pro is particularly competitive in 3D modeling. Engineers can create a 3D model by drawing it or by writing code that specifies its geometry in numerical terms; the model takes the second approach. It scored 92.7% on BenchCAD, a 3D model coding benchmark, placing it within four percentage points of Astra and Claude Opus 5.5. The model can generate static 3D models and animate them. In one Moonshot AI test, it produced an animated simulation of human motion; in another internal evaluation, the company used it to create multiple 3D video games. Games are difficult for AI models because they contain large amounts of code, and a single bug can crash them.

Kimi K3.1 Pro scored 65.7% on Terminal-Bench 4.0, an industry-standard coding benchmark, more than 7% ahead of Astra but slightly behind Opus 5.5. It outperformed both models on OfficeQA Pro v2, a test of complex financial analysis tasks. On popular cybersecurity benchmarks, its performance significantly trails Astra and Opus 5.5. Moonshot AI said its engineers deliberately removed cybersecurity-related information from the model’s post-training dataset to address safety risks, and it plans to equip future models with more advanced vulnerability analysis features.

Kimi K3.1 Pro is available through Moonshot AI’s paid cloud services at launch. The company plans to open-source the model’s weights by Nov. 4.