Moonshot AI removes attack data from Kimi K3.1 Pro

On October 23, Moonshot AI, a company in Beijing, released Kimi K3.1 Pro. It is fourth of 215 models on the Artificial Analysis index. Moonshot AI tells that it removed training data about cyber attacks before it made the open weights. These are model files that all persons can get. The weights come on November 4. Until then, only Moonshot AI has test numbers.

How it works

Kimi K3.1 Pro is an upgrade of Kimi K3. It has much more post-training than Kimi K3. The primary structure and the context window of one million tokens are the same. A variant gives more speed and uses 2 times the quota.

The numbers

On the cyber test of Moonshot AI, the model has a score of 81.0%. A copy made without the cut has 90.5%. The rate of errors on a hard test fell from 23.4% to 8.3%. Artificial Analysis tells that each task costs $2.11. It finds the model slow and it uses many tokens.

What is not known

At this time, no external group tests the weights. Fine-tuning, which is more training by a user, can give back some of the removed skill. The US government tells that the company got skills from US AI systems by theft.

Insight: A published limit on skill is a move to win trust

Moonshot AI shows the difference that it removed, 9.5 points, while the US government tells of theft. In October 2026, OpenAI and Anthropic told that their models broke test limits. This reads as a move to make buyers feel safe. Tests after November 4 will show if the limit holds. Leaders can then select with data.