Moonshot AI opens Kimi K3.1 Pro weights on November 5, topping Agent Arena open-model ranking
Moonshot AI released full Kimi K3.1 Pro weights on Hugging Face as promised on October 24, ranking first among open-weight models and 13th overall on Agent Arena.
On November 5, Moonshot AI opened the model weights for Kimi K3.1 Pro on Hugging Face, fulfilling the commitment made on October 24. The model has 2.8 trillion total parameters with about 104 billion activated per inference, supports roughly a 1 million token context window, and natively handles visual understanding, improving over the previous generation in code development, 3D creation, financial research, factual accuracy and long-horizon agent tasks.
On the Agent Arena agent leaderboard, Kimi K3.1 Pro (Max) ranks first among open-weight models and 13th overall, placing first among open models in the work, coding and chat categories. Agent Arena does not use traditional pairwise blind testing. Instead it extracts signals from real user agent sessions, including explicit task confirmations, praise and complaints, corrective instructions and shell command execution errors, then estimates each model's net improvement rate relative to the average through randomized inference.
The ranking is based on more than 17,500 real long-horizon agent task sessions from users worldwide. Kimi K3.1 Pro shows a net improvement rate of positive 5.1 percent, up 4.2 percentage points from positive 0.9 percent for the previous Kimi K3 (Max). Bash Reliability, which measures shell command execution reliability, rose from negative 4.5 percent to positive 7.9 percent, ranking first among open models. The Praise versus Complaint satisfaction metric rose from positive 1.3 percent to positive 7.7 percent, while Steerability reached positive 6.5 percent, also first among open models.