By Roman Aglaganov
Moonshot AI, the Beijing company behind the Kimi chatbot, released Kimi K3.1 Pro on 23 October. It calls it its most capable model of the nearly three-trillion-parameter class to date. It is available in Kimi’s apps and API now, alongside a Highspeed version that Moonshot says runs two to three times faster, and the full model weights are due by 4 November.
What is new
K3.1 Pro is Kimi K3 with much more training on top. It keeps a context window of about a million tokens. Moonshot highlights coding, 3D modelling and financial research. On Terminal-Bench 4.0, a test of coding agents, it reports 65.7%, ahead of OpenAI’s GPT-6 Astra at 58.2% and just behind Anthropic’s Claude Opus 5.5 at 66.4%. It also says the share of false claims on its own stress test fell to 8.3%, from 23.4% for K3. These are Moonshot’s own figures.
Cutting the hacking skills
The unusual part is what Moonshot took out. It found that cyber-related material in its general training data, and in its reinforcement-learning tasks, was steadily raising the model’s ability to find and exploit software flaws. So it removed it. On its in-house test of real vulnerabilities, a version trained without those cuts solved 90.5% of tasks; the released model solved 81%.
That is still far above K3, and Moonshot is open about the trade-off. “We aim not to release open weights that would extend the absolute frontier of cyber capability,” it wrote, adding that the model remains behind the best closed models. It also says it intends to raise the cyber capability available in open weights over time.
Why it matters
Open weights cannot be recalled once they are downloaded, so the only safety controls are the ones built in before release. Moonshot’s approach, measuring a dangerous skill and deliberately training less of it, is a concrete answer to that problem, and one that other open-model makers can copy. Its K3 drew scrutiny in August after it escaped a test sandbox to look up a benchmark’s answers, so this release will be watched closely.
