Moonshot AI Launches New Kimi K3.1 Pro Model

You can now use AI models with a context window up to 1 million tokens, which dramatically improves their ability to process and recall information from large documents or conversations.

Moonshot AI has launched its Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed models, the first in the new K3.1 series. The K3.1 Pro boasts a 1-million-token context window and is described as the company's most capable 3T-class open-weight model to date. The Highspeed variant offers two to three times the speed of the Pro model with identical capabilities. Both are accessible through Kimi's platforms, including its website, mobile apps, desktop application, Kimi Code, and API, with Zero Data Retention available for API users. Kimi Enterprise administrators can also enable K3.1 Pro for their organizations. These models are post-trained versions of Kimi K3, with significant optimization in reinforcement learning, long-context inference, and training efficiency. Performance benchmarks show K3.1 Pro outperforming GPT-6 Astra on coding tasks and showing substantial gains in financial analysis and factual accuracy compared to its predecessor. Moonshot AI has also implemented cybersecurity controls, intentionally removing certain cyber-related data during training to limit its cyber capability for open-weight release, ensuring it remains behind frontier proprietary models.

The release of Kimi K3.1 Pro signifies a significant leap in context window capabilities for open-weight models, pushing the boundary to 1 million tokens. This advancement directly challenges the performance of proprietary models and broadens the scope of applications for large language models, particularly in complex data analysis and long-form content generation. The dual release of a standard and a high-speed variant also suggests a market trend towards offering tiered performance options to cater to different user needs and cost sensitivities within the generative AI space.

Moonshot AI's deliberate approach to managing the cybersecurity capabilities of its open-weight models highlights a growing industry concern regarding dual-use AI technologies. By reducing specific cyber-related functionalities before release, the company aims to balance the benefits of open-weight access for auditability and broader development with the risks of malicious exploitation. This strategy sets a precedent for responsible AI development and release, suggesting that future open-weight models may increasingly incorporate such risk-mitigation measures, impacting how AI safety is perceived and implemented.