software · DiggingBeagle record
Kimi K3.1 Pro
Moonshot model with reported long-horizon coding capability and a published offensive-cyber ablation.
- Entity kind
- software
Record description
Moonshot AI model announced in October 2026, with approximately three trillion parameters and a one-million-token context. Vendor-hosted variants were available at announcement; a full open-weight publication was planned for November 4 and must not be described as already completed. Its disclosed cyber-training ablation is an unusually useful paired measurement, but the offensive benchmark is vendor-controlled.
The published xdev-bench-hard pass@3 ablation reports 90.5% for an unrestricted K3.1 training variant, 81.0% for the risk-reduced K3.1 Pro and 33.3% for predecessor K3. The 9.5 percentage-point restricted-versus-unrestricted difference is on a vendor-selected offensive exploit benchmark, not a proof that the released model cannot help with an intrusion.