Loki’s Lab Puts Local AI Models Through Their Paces

Field notes

The leaderboard uses fleet-skill-matrix v2 and groups only like suite versions together. Each model is tested three times on a fixed Hermes profile, with incapable runs counted as zero and speed reported as the median across applicable runs. The visible table includes hardware context—such as a Mac mini with an M2 Pro, 16GB memory, and macOS 15—alongside coverage counts and verification states. Community entries require JSON validation, privacy clearance, and explicit approval.