Series
State of the local LLM
Monthly snapshots of what the fastest local LLM actually is, measured under one reproducible workload suite (llm-speed suite-v1). Each issue answers the question “as of YYYY-MM, the canonical answer is X” with deltas vs. the previous month and a citation per claim.
Issues
The leaderboard passes 200 signed runs, the large majority measured this month across RTX 5090, RTX 4090, and Apple Silicon. The fastest decode, the small mixture-of-experts coders that lead the mid tier, and where Apple wins on memory.
Inaugural issue. Fastest local 70B-class result, fastest dense Apple Silicon decode, and the first numbers under suite-v1 with the dual-domain trust chain in place.
Want the next issue in your inbox? There is no inbox yet — subscribe to the public RSS feed (coming soon) or follow github.com/meadow-kun/llm-speed for releases.