Kimi K3.1 Pro Weights Released Open Source; Open Model Enters Agent Arena Overall Top 15 for the First Time

Kimi K3.1 Pro's weights are out, and you can download them directly on HuggingFace.

On the same day, Arena.ai updated the Agent Arena leaderboard: Kimi K3.1 Pro (Max) ranks first among open models and 13th overall. The evaluation is based on 17.5K+ long-horizon agent sessions from real users, with a net improvement of +5.1%.

The number doesn't look large on its own. But when you break it down by category, the interesting part is inside.

It takes first place among open models in all three categories: Work, Code, and Chat. Overall, it is +4.2 percentage points above the previous generation, Kimi K3 (Max).

Bash Reliability (the rate of successful shell command execution) went from 13th place and -4.5% for K3 to first among open models and +7.9%. Steerability (responding to user corrections) rose from +0.4% to +6.5%, first among open models. Praise vs Complaint (user sentiment feedback) went from +1.3% to +7.7%, also first among open models. Confirmed Success (users explicitly saying "it worked") is +3.9%, fourth among open models.

One commenter wrote: a 5-point improvement shows they finally fixed the habit of "hallucinating nonexistent CLI flags."

Another comment is shorter: Listening when corrected is rarer than topping a chart. Being able to take corrections is rarer than reaching the top.

Agent Arena measures long-horizon tasks, not single-turn Q&A. The user makes a request, and the model executes multiple steps in a row, getting corrected, interrupted, and asked to start over along the way. The Bash Reliability and Steerability sub-scores map exactly to this kind of scenario. Companies that just chase leaderboard rankings are unlikely to put much effort into these two areas.

Databricks launched Kimi K3.1 Pro the same day. Engineer Yuchen Jin said the team is already using it internally to write code. PerComputing said it runs at 200 tokens/s on their platform.

Weights address: https://huggingface.co/moonshotai/Kimi-K3.1-Pro

In the comments, someone wrote, "download it before anyone changes their mind," someone is waiting for the community to release a quantized version, and someone said they're already running it on a Mac Mini. There is also a reply image captioned "someone check on dario."

At this stage of competition among open models, releasing the weights is only the starting line. Whether a model can stay stable in real agent tasks, take advice, and make fewer dumb mistakes is what to watch in the next round.