Moonshot AI has launched Kimi K3.1 Pro, enhancing programming and 3D modeling while limiting cybersecurity capabilities.
Chinese tech firm Moonshot AI has unveiled its latest language model, Kimi K3.1 Pro, which boasts enhanced programming, 3D modeling, and financial analysis capabilities. However, the developers have intentionally restricted its ability to search for and exploit vulnerabilities.
Kimi K3.1 Pro is an upgraded version of the Kimi K3 model, built on a Mixture-of-Experts (MoE) architecture comprising 2.8 trillion parameters, of which 104 billion are activated during query processing. It features a context window of 1 million tokens and can process both text and images.
Alongside this, Moonshot AI has also introduced Kimi K3.1 Pro Highspeed, which reportedly generates responses two to three times faster than the standard version. Both models are accessible via the Kimi website, applications, and APIs, with full model weights expected to be released by November 4.
Performance Metrics
In the Terminal-Bench 4.0 test, which evaluates AI agents on engineering tasks, Kimi K3.1 Pro achieved a score of 65.7%. For comparison, Claude Opus 5.5 scored 66.4%, and GPT-6 Astra scored 58.2%. However, the testing methodologies varied, as Moonshot AI used its own Claude Code for testing while competitor scores were based on published reports.
Within the Kimi Code Bench 3.0 Long-horizon, which assesses prolonged programming tasks, K3.1 Pro scored 61.2%, a notable improvement over its predecessor's 55.95%. Independent evaluations by Artificial Analysis also confirmed progress, with the new model scoring 55 in the Intelligence Index compared to 44 for the earlier version. K3.1 Pro produced 14% more output tokens on average while maintaining API task costs similar to K3, despite a one-third reduction in API rates.
3D Modeling and Financial Analysis Features
Kimi K3.1 Pro has enhanced capabilities for 3D graphics, enabling the creation of ready-to-use 3D models from text descriptions, sketches, and images, accurately determining shapes, proportions, and arrangements. The system can autonomously write and execute scripts for Blender and CAD programs, modifying geometry, materials, and layouts for various applications such as mechanical design and architectural structures.
In the BenchCAD test for reconstructing 3D objects from images, K3.1 Pro scored 92.7 in Voxel IoU metrics, while Claude Opus 5.5 scored 96.2 and GPT-6 Astra 95.9. The model also excels in financial analysis, capable of processing corporate reports, correlating data across documents, and constructing financial models.
In OfficeQA Pro v2, which evaluates financial document analysis and multi-step calculations, K3.1 Pro scored 64.4, with improvements noted in financial modeling and investment research compared to K3.
Cybersecurity Measures Implemented
Moonshot AI has placed significant emphasis on the model's security features ahead of the planned release of its weights. During development, engineers observed that the model's ability to identify and exploit vulnerabilities increased with enhanced training, prompting the decision to exclude cybersecurity-related data from its training corpus.
In an internal benchmark, a test version without these restrictions succeeded in 90.5% of tasks, while the final K3.1 Pro achieved 81%. The company acknowledges that even with these limitations, the new model demonstrates substantial programming skill improvements, although it still lags behind leading proprietary models available for at least three months.
Moonshot AI clarified that it opted not to release the model's open weights to prevent potential misuse in cyberattacks while highlighting the utility of such tools for software security audits. The API pricing for standard K3.1 Pro is set at $2 per million incoming tokens and $10 per million outgoing tokens.
This launch follows the introduction of Kimi K3 in July, which was touted as the first open model of its kind with native image support, designed for long-term autonomous programming tasks.
