Moonshot AI has released Kimi K3.1 Pro, a large language model that it says outperforms some leading models on benchmarks, while the company says it removed cyber-related training data to address safety risks.

• Moonshot says it removed cyber-related data from the model's mid-training and reinforcement learning stages, and the released model solved 81.0% of its internal exploit test versus 90.5% for an unfiltered build.
• The company says it plans to open-source the model's weights by November 4, and that the model uses a hybrid linear-attention mechanism called Kimi Delta Attention to reduce memory use.
• Two Moonshot documents give different hallucination rates for K3.1 Pro: 8.3% in the blog and 6.85% in the API quickstart, and neither refers to the other.