China’s Moonshot AI unveils 3-trillion-parameter Kimi K3.1 Pro

Chinese AI company Moonshot AI has introduced the Kimi K3.1 Pro, a groundbreaking 3-trillion-parameter open-weight model with an impressive 1-million-token context window. The company announced the release of both Kimi K3.1 Pro and its faster counterpart, Kimi K3.1 Pro Highspeed, through various platforms including Kimi.ai, mobile apps, Kimi Work, Kimi Code, and the Kimi API.

Moonshot AI highlighted that the model boasts significant improvements across various domains such as coding, 3D creation, financial research, and other agentic capabilities.

Developed with enhanced computing resources during post-training, the model aims to deliver substantial performance across hundreds of application areas. Moonshot AI emphasized that the model supports a broad range of coding tasks, from fixing existing codebases to creating complete projects, as well as offering end-to-end financial workflows.

The AI company optimized its reinforcement learning framework, long-context processing, training efficiency, and infrastructure to enhance performance across various application domains.

Kimi K3.1 Pro has demonstrated competitive performance on several benchmarks. According to Moonshot AI, the model scored 65.7 percent on Terminal-Bench 4.0, a metric evaluating AI coding agents on real-world terminal-based tasks, outperforming GPT-6 Astra's 58.2 percent and coming close to Claude Opus 5.5's 66.4 percent. In the internal Kimi Code Bench 3.0 Long-horizon benchmark, K3.1 Pro achieved a score of 61.2 percent, compared to 55.95 percent for the previous K3 model.

Additionally, Moonshot AI introduced Kimi K3.1 Pro Highspeed, which the company claims delivers two to three times faster performance than the standard K3.1 Pro model for swifter interactions.