SiliconANGLE reported on Oct. 23 that China's AI company Moonshot AI has unveiled its new flagship large language model, Kimi K3.1 Pro. The model outperformed Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra on some tasks.

Kimi K3.1 Pro is an upgraded version of Kimi K3 released in July, improved through post-training, the report said.

Kimi K3 is a model with 3 trillion parameters and is made up of 896 smaller neural networks. When it receives a query, it selects 16 of them to do the work. Because it does not run the entire system, it can reduce operating costs.

Moonshot AI put far more computing resources into this post-training than before. It added automated quality checks to the training stack and improved infrastructure efficiency.

The new model is particularly strong in 3D modelling.

In the 3D model coding benchmark BenchCAD, it scored 92.7 percent, narrowing the gap with Astra and Opus 5.5 to within 4 percentage points.

In the coding benchmark TerminalBench 4.0, it scored 65.7 percent. That was more than 7 percent higher than Astra and slightly lower than Opus 5.5. In OfficeQA Pro v2, which evaluates financial analysis ability, it outperformed both models.

Its cybersecurity benchmark score fell well short of both models. Moonshot AI explained it deliberately removed some cybersecurity information from training data due to safety risks. It plans to strengthen vulnerability analysis features in its next models.

Kimi K3.1 Pro is available immediately through Moonshot AI's paid cloud service. Moonshot AI plans to release the model weights as open source until Nov. 4.