Kimi unveils K3.1 Pro, a 3T-class open-weight model with a 1M-token context window, and K3.1 Pro Highspeed, with full weights to be released by November 4

Open-weight models now match or approach proprietary AI performance on key benchmarks.

Kimi has released Kimi K3.1 Pro, a new 3T-class open-weight language model featuring a 1-million-token context window. This model offers significant enhancements in coding, 3D creation, and financial research capabilities, performing comparably to leading proprietary models on various benchmarks. For instance, it scored 65.7% on Terminal-Bench 4.0 for coding agents, outperforming GPT-6 Astra. Kimi also introduced Kimi K3.1 Pro Highspeed, a faster version for more responsive interactions. Both models are available now via Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, with full model weights scheduled for release by November 4th. Kimi K3.1 Pro is a post-trained version of Kimi K3, benefiting from extensive post-training computation to improve its performance across numerous domains.

Kimi's release of K3.1 Pro signals a significant advancement in open-weight models, demonstrating that these models can now compete with or even surpass top-tier proprietary offerings in complex tasks like coding and financial analysis. The substantial 1-million-token context window allows for processing and reasoning over much larger amounts of information, which is critical for intricate, long-horizon tasks previously out of reach for many models. This development suggests a potential democratization of advanced AI capabilities, enabling a broader range of developers and organizations to build sophisticated applications without relying solely on closed commercial systems.

The implications of K3.1 Pro's performance, especially its strong showing on coding and 3D generation benchmarks, indicate a shift in the ecosystem. Companies that previously relied on proprietary models for cutting-edge development may now find viable open-weight alternatives, potentially lowering costs and increasing flexibility. Furthermore, the inclusion of risk mitigation measures, such as removing cybersecurity-related data, highlights a growing industry awareness and proactive approach to responsible AI deployment, which could set new standards for the development of open-source AI models.