Projects / Benchmarks
Benchmarks
A score is never just the model's.
An open leaderboard for local and hosted inference. Model, version and engine are measured with one harness held constant, so a difference in scores can be traced to its cause.
What it is
Each score is tied to three things: the model family and size, the exact checkpoint and quantization, and the engine with its runtime, flags and hardware.
One harness, bench, runs every test.
The categories are Language, Decision, Embedding and Image. As of the latest run on 2026-10-09, the board holds 51 runs across 33 subjects.