Moonshot AI releases Kimi K3.1 Pro weights, topping the open agent leaderboard

Moonshot AI has released the model weights of Kimi K3.1 Pro on Hugging Face, keeping its earlier promise. The model uses a mixture-of-experts design with 2.8 trillion total parameters, about 104 billion active per run, and supports a context window of around 1 million tokens. On the Agent Arena leaderboard, Kimi K3.1 Pro ranks first among open-weight models and 13th among all models. Its net improvement rate is +5.1%, up 4.2 points from the previous K3's +0.9%, based on over 17,500 real agent task sessions from users worldwide. On the Terminal-Bench 4.0 agent test, it scored 65.7%, beating GPT-6 Astra's 58.2% and just under 1 point behind Claude Opus 5.5's 66.4%. However, the BF16 weights take up about 5.6TB, and even INT4 quantization needs around 1.4TB, so it is hard to run locally on consumer devices.