Moonshot AI releases 2.8 trillion parameter kimi K3.1 pro open weights model, tops agent arena leaderboard

Moonshot AI made the Kimi K3.1 Pro model weights available on Hugging Face today, fulfilling a commitment made on October 24. The model achieved the top spot among open-weight models on the Agent Arena leaderboard, ranking 13th overall across all evaluated models, and leading in the categories of work, programming, and chat. This ranking is based on over 17,500 real-world long-cycle agent tasks, where Kimi K3.1 Pro showed a net improvement rate of +5.1%, a 4.2 percentage point increase over the previous K3 (Max) model's +0.9%. Specific gains include a Bash Reliability score of +7.9% (ranking first among open models) and a Praise vs Complaint score of +7.7%. Technically, Kimi K3.1 Pro utilizes a Mixture-of-Experts architecture with 2.8 trillion parameters and supports a 1 million Token context window. While the model performs strongly, requiring about 1.4TB even with INT4 quantization, its size presents significant deployment hurdles for consumer-grade hardware. Furthermore, on the Terminal-Bench 4.0 benchmark, K3.1 Pro scored 65.7%, narrowly trailing Claude Opus 5.5 (66.4%) but surpassing GPT-6 Astra (58.2%).