Moonshot unveils Kimi 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 Nov. 4

Open-weight models now match or exceed proprietary AI in specialized tasks like coding and financial analysis.

Moonshot AI has launched Kimi K3.1 Pro, an open-weight model boasting a 3 trillion parameter class and an unprecedented 1 million token context window. This new model demonstrates significant advancements in coding, 3D creation, financial research, and agentic tasks, with performance benchmarks positioning it competitively against leading proprietary models like GPT-6 Astra and Claude Opus 5.5. A faster version, Kimi K3.1 Pro Highspeed, is also available, offering 2-3x improved responsiveness. While Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed are accessible now via Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, the full model weights are scheduled for release by November 4th. Moonshot AI emphasized optimized training and inference infrastructure to achieve these gains, particularly in long-horizon tasks and complex reasoning.

The release of Kimi K3.1 Pro signals a maturing landscape for open-weight large language models, particularly in their ability to compete with closed-source alternatives on complex benchmarks. Moonshot AI's achievement in integrating a 1-million-token context window with a 3T-class model, while also optimizing for speed and specific agentic capabilities, suggests that open models are rapidly closing the gap in areas requiring extensive context or intricate multi-step reasoning. This could accelerate adoption of advanced AI capabilities in industries that prioritize transparency, customization, or cost-effectiveness.

The strategic removal of cybersecurity-related data during Kimi K3.1 Pro's training highlights a critical ongoing challenge for AI developers: balancing model capability with safety and misuse prevention. By demonstrating a method to identify and restrict specific data types without sacrificing general performance, Moonshot AI offers a potential blueprint for responsible AI development. This approach could influence how future models are trained and evaluated, potentially leading to increased scrutiny of training datasets for embedded risks across various domains, not just cybersecurity.