Moonshot says Kimi K3.1 Pro scored lower on its internal vulnerability test after cyber-related material was removed. Reuters reported that the results have not been independently verified.

Moonshot rolled out Kimi K3.1 Pro in Kimi Code on October 23. It plans to publish the model weights on November 4. The key question is whether the company’s reported risk reduction will hold when developers can run and modify the model themselves.

For now, K3.1 Pro is available through Kimi Code. Open weights would give developers more control to run the model on their own systems. But, as Reuters reported, Moonshot would have less control over access and later changes than it does with a hosted service.

What Moonshot says it changed

Moonshot says it removed cyber-related data from the model’s training corpus and reinforcement-learning environments while keeping its general coding skills. The company compared the model’s performance on an internal test of finding and exploiting software vulnerabilities:

  • With the risk-reduction measures: 81.0%.
  • Without those measures: 90.5%.
  • Difference: 9.5 percentage points.

The gap is a result from Moonshot’s own test, not independent proof that the model is safe from misuse. Reuters reported that the results have not been independently verified. The 81.0% score also shows the model retained substantial performance on that particular test after the changes.

Hosted access comes before the weights

Kimi Code’s release notes say K3.1 Pro rolled out on October 23 with a one-million-token context window for eligible plans. Moonshot also offers K3.1 Pro Highspeed, which it says runs about two to three times faster. The largest context window and Highspeed option require higher-tier plans.

The planned November 4 release would change who controls access. Hosted users rely on Moonshot’s service. Developers with the weights could run and modify the model on their own systems. That flexibility can help organizations inspect software, but it also leaves Moonshot with less ability to monitor use or update safeguards after the files spread.

Moonshot’s figures compare its own training choices. They do not establish how the model will be used once the weights are available.