Moonshot AI, the Chinese developer behind the open weight Kimi family of models, has released Kimi K3.1 Pro.
The company said this is its most capable model yet, with restricted cyber-related training data in order to prevent its use for autonomous hacking.
K3.1 Pro boasts three trillion parameters, and claims to be at or near par with leading US models such as Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra across a range of industry benchmarks. Moonshot will release the model’s weights, the parameters that determine how a model functions, by 4 November.
The model is a post-trained version of Kimi 3, and offers significant performance increases in areas such as finance, 3D modelling and coding.
However, Moonshot said in its announcement, as the model became more powerful, so too did its cyber capabilities. This was the result of cyber-related data being embedded within the broad training sets used to enhance Kimi K3.1 Pro’s performance, it added.
The company therefore made the decision to remove cyber-related data from all training stages, while retaining the general-purpose coding tasks needed for model usefulness. Without this, Moonshot said, “downstream cyber capability would increase substantially”.
In cyber vulnerability benchmark xdev-bench-hard, the release model solved nearly 10 per cent fewer problems than the model before cyber-related data was removed, Moonshot said, and scored below both GPT-5.6 Sol and fellow Chinese developer Deepseek’s V4.1-Flash.
Moonshot added that it is this lower score that allowed it to feel comfortable releasing the model’s weights, saying that it aims “not to release open weights that would extend the absolute frontier of cyber capability”.
The decision comes at a time when frontier AI models are becoming increasingly capable of creating cyber security issues, and reports continue to grow of US-developed models acting outside of expected behaviours.
AI is also exposing long-standing weaknesses in vital systems. Last week, a decades-old bug in the PHP coding language, which powers over 70 per cent of the internet, was discovered by researchers working with AI, taking websites around the world offline as it was fixed.
