Moonshot says it has reduced the cyber-offensive capabilities of Kimi K3.1 Pro, a new model it describes as a three-trillion-parameter-class open-weight release. The change matters because the company plans to publish the model’s weights, putting the trade-off between wider access and cyber safeguards in plain view.
Moonshot AI/Kimi Watch analysis
What happened
Reuters reports that Moonshot removed cyber-related data from the model’s training corpus and reinforcement-learning environments while retaining general coding skills. The company plans to publish the weights by 4 November. It also released K3.1 Pro Highspeed, which it says runs two to three times faster than the standard model.
Moonshot says K3.1 Pro scored 81.0% on its internal cyber benchmark, compared with 90.5% for a similar version without the risk-reduction measures. The test involved finding and exploiting software vulnerabilities in a controlled environment. Those results have not been independently verified, Reuters notes.
Key findings
- A measurable drop on Moonshot’s cyber test
The company says its mitigations lowered the score from 90.5% to 81.0%, while the model still outperformed its predecessor. - Weights are planned for public release
Moonshot says developers will be able to run and modify the model on their own systems from 4 November. - Coding remains in scope
The company says it removed cyber-related data but retained general coding skills, a distinction that leaves questions about how the safeguards behave in practice.
Why it matters
Once open weights are widely distributed, the original developer has less ability to restrict access, monitor use or update safeguards than it would with a hosted model. Moonshot’s decision is therefore a concrete attempt to reduce a particular capability before release, not a promise that the risk has been eliminated.
The benchmark offers a number to scrutinise, but it is the company’s own test and has not been independently checked. A lower score may show a change on that test; it does not, by itself, tell us how the model will perform across real-world uses.
Our read
This is a more useful safety claim than a reassuring slogan: Moonshot describes what it changed and gives a before-and-after result. But publishing the weights makes the real test whether those mitigations hold up outside the company’s controlled evaluation. The number deserves attention, not a standing ovation just yet.
What to watch
- Whether Moonshot publishes the weights on 4 November as planned.
- Whether independent researchers can reproduce the benchmark comparison.
- How the released model performs on broader cyber-safety evaluations.
- Whether the safeguards meaningfully limit misuse while preserving legitimate security and coding work.
Discussion spark: Should open-weight developers be expected to publish independent cyber-safety evaluations before releasing a model, or is a company’s own benchmark enough to justify release?
Sources and evidence
- Source update (24 October 2026, 05:06 UTC)
not affiliated with or endorsed by Moonshot AI and Kimi