Overview
- Moonshot introduced Kimi K3.1 Pro this week, saying the model uses a mixture-of-experts design with about 2.8 trillion total parameters, roughly 104 billion active parameters at a time, and a native one-million-token context window.
- The company published internal benchmarks that claim strong performance on coding and 3D tasks, including 65.7% on Terminal Bench 4.0 and 92.7% on BenchCAD, and reported a drop in hallucinations from 23.4% to 8.3% on a hard internal test.
- Moonshot said it removed cyber-related training data and tasks during development and reported that the filtered K3.1 Pro scored 81.0% on its internal exploit benchmark versus 90.5% for an unfiltered version, a measure the company presents as a safety trade-off.
- Moonshot will publish full model weights on November 4, which will let outside groups run and modify the model locally and enable independent checks of the company’s capability and safety claims.
- The release sharpens a policy debate because open weights can speed defensive research while also lowering the developer’s ability to control misuse, and regulators including Chinese authorities are weighing export limits on AI tools that help generate malware or find vulnerabilities.