Chinese AI lab Moonshot has released Kimi K3.1 Pro, and says it stripped hacking-related material from the model's training data to limit misuse.
The model is an upgrade to Kimi K3, which is open-weight: the trained model files can be downloaded, run and modified by anyone, rather than reached only through a company's servers. Moonshot said the full weights are due by November 4. The company also said K3.1 Pro nearly matches Claude Opus 5.5 on a coding test.
Both claims come from Moonshot, and neither comes with much detail. The company has not disclosed which coding test it used, the scores, how much data it removed, or how it defines "hacking-related." Until it does, "nearly matches" is a marketing phrase, not a measurement.
The safety approach is the more interesting part. Most AI labs limit misuse with refusal training, which teaches a model to decline certain requests. That works reasonably well on a hosted service, where the lab controls the model. It is a weaker fit for open weights, because anyone with the files can fine-tune a model to change how it behaves. Filtering the training data works at a different stage: it targets what the model learned, not what it has been told to decline.
That approach has a cost question attached. Security knowledge overlaps heavily with ordinary programming, and defenders use models to review code and understand attacks. A filter broad enough to matter could also dull the coding skills Moonshot is promoting, and one narrow enough to protect them may leave gaps.
The November 4 weights release is when outside testers can start checking both sides. For anyone weighing an open model for real work, the useful numbers will be how it scores on coding benchmarks and how it handles security tasks, not what the launch post says about either.