# /compare benchmark data (2026-09)

- `run2/` — M3 Pro 18 GB, all engines, Rapid-MLX 0.15.2 (its versions.json is `raw/versions.json`). Competitor numbers for the 18 GB table.
- `run3-m4pro-48gb/` — M4 Pro 48 GB, all engines, Rapid-MLX 0.15.2. Competitor numbers for the primary (48 GB) table; its Rapid-MLX rows are the before-fix baseline.
- `run4-m4pro-48gb/`, `run4-m3pro-18gb/` — Rapid-MLX only, measured at commit 1394f16960d51e5d619c1aa45f1aa173454eb8c8 (includes #3794), which Rapid-MLX 0.15.3 includes. The benchmark was not re-run on the release tag. The competitors were not re-run: same machines, same versions as run3 / run2.
- `run1/` — superseded first run (see the /compare page for why).

Every run uses `bench_compare.py` + `run_all.sh` from the parent directory (run 1: `run1/harness/`). Each run's `versions.json` records machine, versions, commands and notes. `../summary.json` (48 GB) and `../summary-m3pro-18gb.json` (18 GB) are what the pages show; they are generated by `landing/scripts/build_compare_pages.py` from these files.