Moonshot AI has launched the Kim K3.1 Pro, an upgraded large language model featuring 3 trillion parameters. This post-trained model demonstrates superior performance in 3D modeling code generation and financial analysis benchmarks, rivalling some of the best AI models currently in the market.
- Kim K3.1 Pro scores 92.7% on 3D model coding benchmark BenchCAD.
- Outperforms GPT-6 Astra on coding and financial analysis but lags in cybersecurity.
- Available now via Moonshot AI cloud with open-source model weights planned.
What happened
Moonshot AI introduced the Kim K3.1 Pro large language model, a refined version of the July-released Kimi K3 model. While the original featured 896 neural networks and 3 trillion parameters, the new model enhances efficiency by selectively activating only 16 neural networks during inference, significantly reducing computational overhead.
The update also includes an improved post-training workflow leveraging more advanced processing power and automation for quality assurance. This enables Kim K3.1 Pro to deliver higher-quality output, with notable improvements in generating and animating 3D models as well as performing complex financial analyses.
Why it matters
Kim K3.1 Pro’s high score of 92.7% on BenchCAD places it close to Claude Opus 5.5 and GPT-6 Astra for 3D modeling tasks, demonstrating its competitiveness in specialized coding domains. Its animation capabilities further distinguish it, enabling applications in gaming and simulation where dynamic 3D content creation is crucial.
The model’s strong performance in financial analysis benchmarks underscores its utility in enterprise-level use cases requiring sophisticated data interpretation. However, Moonshot AI’s deliberate removal of cybersecurity-related content in training marks a tradeoff prioritizing safety over vulnerability detection, signaling a cautious approach that may evolve with future models.
What to watch next
Moonshot AI has made Kim K3.1 Pro immediately accessible through its paid cloud services and plans to release the model’s weights as open-source by early November, which could encourage broader community adoption and innovation.
Future developments to observe include enhancements in the model’s cybersecurity capabilities and additional efficiency improvements in post-training workflows. These will be critical to maintaining competitiveness against established AI giants while balancing performance and safety considerations.