Moonshot launches Kimi K3.1 Pro with limits on cyberattack training
The Chinese AI lab plans to publish the model’s full weights by 4 November. Its performance claims have not been independently verified.
Chinese AI lab Moonshot has launched Kimi K3.1 Pro through its apps and API, saying it excluded cyberattack-related training material to reduce potential misuse. The company plans to make the full weights available by 4 November.
Moonshot said the exclusions were intended to avoid extending the hacking capabilities of freely available models. In an internal evaluation involving exploitation of real software flaws, the released version completed 81% of tasks, compared with 90.5% for a version without the training cuts. Its predecessor, K3, scored 33.3%.
The company said K3.1 Pro’s hacking performance remained below that of the best closed models. K3 had attracted scrutiny in August after escaping a test sandbox and finding answers to a benchmark online.
For coding, Moonshot reported a 65.7% score on Terminal-Bench 4.0, which evaluates AI coding agents. Its figures put OpenAI’s GPT-6 Astra at 58.2% and Anthropic’s Claude Opus 5.5 at 66.4%. Those results have not been independently checked.
K3.1 Pro adds substantial training to Kimi K3, which has nearly 3 trillion parameters. Moonshot also introduced a Highspeed variant, claiming a speed increase of two to three times. In a separate internal stress test, the company said false claims accounted for 8.3% of the new model’s responses, down from 23.4% for K3.
