November 5 news: As promised at its October 24 release, Moonshot AI today officially fully opened the model weights of Kimi K3.1Pro on the Hugging Face platform. With this official open-sourcing of the weights, the model has performed impressively on Agent Arena, the agent leaderboard operated by Arena, taking first place among open-weight models, ranking 13th among all evaluated models, and holding the top spot among open models in all three core categories: work, coding, and chat.
Rigorous evaluation based on real long-horizon sessions
Unlike traditional pairwise blind tests, Agent Arena uses a unique evaluation method. It does not rely on users voting on two responses; instead, it digs into real users' agent sessions and automatically extracts multiple key signals, including explicit user confirmation of tasks, praise and complaints, corrective instructions, and Shell command execution errors. On this basis, the leaderboard uses a randomized inference method called "causal tracing" to scientifically estimate each model's net improvement rate relative to the average.
Kimi K3.1Pro's impressive ranking this time is based on more than 17,500 real-world long-horizon agent task sessions from users worldwide. Its overall net improvement rate reached +5.1%, a significant increase of 4.2 percentage points compared with +0.9% for the previous-generation K3 (Max).
Looking at specific sub-metrics, all core capabilities have undergone comprehensive improvements:
Bash Reliability (Shell command execution reliability): Soared from -4.5% in the previous generation (ranked 13th) to +7.9%, jumping to first place among all open models.
Praise vs Complaint (user satisfaction metric): Rose steadily from +1.3% to +7.7%.
Steerability (ability to adjust direction and correct course): Reached +6.5%, also taking the top spot among open models.
Confirmed Success (task completion confirmation): Reached +3.9%, ranking fourth among open models.
Architectural innovation and extreme parameter design
