Moonshot says it removed cyber-related data from two training stages of Kimi K3.1 Pro to limit offensive capabilities before a planned release of the model weights. Its current research post is dated October 24 in the official index.
On its internal xdev-bench-hard test, a checkpoint without those filters solved 90.5% of tasks; the final model solved 81%. The metric is pass@3, meaning success within three attempts. Tasks provide source code and a bug hint, with testing in a separate sandbox and no hard time or output limits.
Even after filtering, K3.1 Pro substantially outperforms its predecessor on this test, according to Moonshot. The company attributes this to general coding and debugging skills carrying over to cyber tasks. These are manufacturer-reported results. Reuters also covered the measure on October 25 and notes that the test results have not been independently verified.
Moonshot says it aims not to release open weights that advance the highest available cyber capability. It also intends to increase cyber capability in open models gradually, arguing that this helps defensive users. It promises the full weights by November 4. The published comparison does not show how well the filtering would hold up after further training by others.