COMMUNITY BENCHMARKS LIVE

Real numbers from
real Macs.

Measured by the people who own them. Every submission is one locked run — two prompt sizes, five rounds after a discarded warmup, greedy, prefix cache off, one request at a time — which is the only reason two people's numbers can be compared.

rapid-mlx benchmark run + share puts your result in the Community Benchmark beta and on your own contributor page. The ranked board below takes the legacy rapid-mlx bench --submit flow.

Share a run from your Mac See how runs are measured
FASTEST DECODE ON THE BOARD
—tok/s
—
Loading the frontier record…

The board

Every row is one measured cell — one chip, memory size, model and acceleration arm. The board opens on every Mac at once, so those rows are raw observed results; filter to a single Mac and memory size to compare like-for-like.

Comparing

Who's contributing

Each Mac gets a name from a random per-install id — never a serial number. We know the name; we don't know the person.

Anonymous benchmark identities — avatars are chosen by a random install id, not a person. Nothing is stored, and deleting the id file resets it.

Latest submissions

Newest first, straight off contributors' machines.

    Where the board still can't answer

    Outlined means nobody has run it yet, so the board genuinely cannot answer for it. If one of these is yours, you're the only person who can fill it in.

    Macs
    Models measured so far
    Open invitations

    Benchmarks waiting for a first run

    Questions the board cannot answer yet. Where the shipped CLI can produce the missing rows, the command is right there — run it and the answer becomes everyone's. Each card says which flow it uses: the Community Benchmark flow lands in the beta section below, the legacy bench flow lands on the board.

    Beta Visible evidence, not ranked

    Community Benchmark beta

    Recent local-first LLM, image and video runs — this is where rapid-mlx benchmark run + share results appear. Identity is not verified yet, so these are visible evidence — not ranked comparisons, and never mixed into the board above.

    Representative community results

    Latest contributions

    All public beta data

    Share a run from your Mac

    Run locally first. Sharing is a separate command that previews the exact payload and asks for consent before anything leaves your Mac.

    Where it lands: the Community Benchmark beta above, and your contributor page — share prints its address. It is visible evidence, not a ranking, and it does not enter the board; the board takes the legacy rapid-mlx bench <alias> --submit flow.

    Published: chip, RAM, core counts, OS and package versions, the model alias, and the timings.

    Never sent: prompts, files, or any hardware serial. The id behind your board name is random per install and resets when you delete ~/.rapid-mlx/bench-install-id.

    Never stored: your IP — it is used transiently for abuse limits.

    Legacy data: /api/benchmarks Beta data: /api/benchmarks/atomic/public Methodology: performance flags CLI: benchmark reference · bench reference (legacy)