Moonshot AI has released the weights of its Kimi K3.1 Pro model as open source on Hugging Face. The model now ranks first among open-weight models on the Agent Arena agent leaderboard and 13th overall among all evaluated models, taking the top spot among open models in all three categories: work, coding, and chat. The evaluation draws on more than 17,500 real long-horizon agent task sessions, and the model shows a net improvement rate well above that of the previous generation, with strong results across multiple sub-metrics. It uses a mixture-of-experts architecture with 2.8 trillion total parameters, activating roughly 104 billion parameters per inference, and supports a context window of about 1 million tokens. The model was post-trained through large-scale joint reinforcement learning, and its attention layers were optimized to cut the overhead of long-context modeling. Given the sheer size of the weights, deployment on consumer-grade GPUs is impractical. On the Terminal-Bench 4.0 benchmark, it outscores GPT-6 Astra and comes close to the level of Claude Opus 5.5.