Moonshot curbed the cyber capabilities of Kimi K3.1 Pro — The Express Tribune

Chinese startup Moonshot AI said it had limited the capabilities of its new Kimi K3.1 Pro model for offensive cyberattacks. The company said it removed cyber-related materials from the training dataset and reinforcement learning environments while retaining general programming skills.

Model with open weights

As The Express Tribune reports, Moonshot unveiled Kimi K3.1 Pro as its most powerful open-weight model in the 3-trillion-parameter class. The company plans to publish the model's weights — files that will allow third-party developers to run and modify it on their own systems — by November 4.

Moonshot said it does not seek to release open model weights that would push the boundaries of cyber capabilities. The company also believes that open-weight models can help organizations test and protect their own software, including legacy systems that receive little attention.

More current news is available on the UA.News Telegram channel Telegram.

Internal test results

According to Moonshot, K3.1 Pro scored 81.0% on an internal cyber benchmark, while a similar version without risk-reduction measures scored 90.5%. The test involved finding and exploiting software vulnerabilities in a controlled environment.

The company noted that filtering cyber data reduced the model's offensive capabilities, although it still outperformed the previous version. These results have not been independently verified. The model has a context window of 1 million tokens, allowing it to process large volumes of text. Moonshot also introduced a faster K3.1 Pro Highspeed version, which, according to the company, operates two to three times faster than the standard model.