Moonshot AI has officially launched its latest flagship large language model, Kimi K3.1 Pro, marking a significant evolution in the company’s high-performance AI architecture. Featuring a massive 3-trillion parameter count, the new model arrives as an upgraded iteration of the Kimi K3 series, which first debuted in July. By leveraging advanced architectural refinements, the company aims to challenge current industry leaders while maintaining a strategic focus on inference efficiency.
At the core of the K3.1 Pro is a sophisticated neural network design comprising 896 individual sub-networks. To optimize operational costs and hardware utilization, the model employs a sparse activation strategy, engaging only 16 of these networks at any given time. This approach, combined with the proprietary Kimi Delta Attention mechanism—a custom implementation of Gated DeltaNet technology—allows the model to significantly reduce RAM consumption during the interpretation of complex user prompts.
The development of K3.1 Pro required a substantial increase in computational resources compared to previous iterations. Moonshot AI utilized an upgraded post-training stack, incorporating automated quality assurance workflows and enhanced infrastructure efficiency protocols. These improvements have yielded notable results in specialized tasks, particularly within the realm of 3D modeling and simulation. In testing against the BenchCAD benchmark, the model achieved a 92.7% success rate, demonstrating a high aptitude for generating 3D geometry through numerical code rather than manual input.
Beyond static modeling, the K3.1 Pro has shown proficiency in complex animation and game development. The model successfully generated animated simulations of human motion and demonstrated the ability to produce functional code for 3D video games—a task notoriously difficult for AI due to the high sensitivity of game engines to code errors. In the Terminal-Bench 4.0 coding evaluation, the model recorded a 65.7% score, outperforming several prominent industry alternatives in code generation capabilities.
The company also highlighted the model’s performance in the OfficeQA Pro v2 benchmark, where it excelled at complex financial analysis. However, Moonshot AI noted that the model’s performance on cybersecurity benchmarks remains intentionally lower than that of its competitors. Engineers opted to remove specific cybersecurity-related datasets from the post-training phase to mitigate safety and vulnerability risks, with plans to integrate more robust security analysis features in future releases.
Kimi K3.1 Pro is currently accessible through Moonshot AI’s paid cloud service offerings. In a move aimed at fostering broader ecosystem adoption, the company has announced plans to release the model’s weights as open-source by November 4, 2026.
