Real inference numbers from real Macs. The headline benchmarks on the homepage are measured on one M3 Ultra
— this page is how everyone else fills in their own row. Submitted by users via
rapid-mlx bench <alias> --submit, aggregated by (chip × model × Rapid-MLX version),
median across submissions, IQR shown for groups with multiple rows.
| Chip | Model | Version | Short decode tok/s | Long decode tok/s | Long TTFT (ms) | Rows |
|---|---|---|---|---|---|---|
| Loading aggregate… | ||||||
Got an M2 Air, M4 Max, M3 Ultra with different RAM? Run the standardized bench and open a PR — the
schema validator + maintainer review run on GitHub. No data leaves your machine without an explicit
y/N confirmation, and the runner only reads
non-privileged macOS interfaces.
$ rapid-mlx bench <alias> --submit
Source rows:
community-benchmarks/submissions
·
Raw aggregate: aggregated.json
·
Schema + methodology: community-benchmarks/
·
Single-vendor headline numbers: /#performance