What operators are actually running. Every row here was submitted by someone running runlocalai-bench, reviewed by editorial, and gated by reproduction status. Empty cells mean the submitter didn't report that number — never a fabrication.
curl -fsSL https://runlocalai.co/bench.mjs | node --hardware <slug> --model <tag> --submit
5 runs × ~2 min. POSTs to /api/community/benchmarks on success.
Auto-parses 9 fields (model, cold-start, median, P5/P95, variance, OS, runtime, version, runs). Pick the matching hardware. Submit. 30 seconds total.
Open paste mode →Both paths run the same protocol, hit the same review queue, and land here once approved. Anonymous OK. Source on /guides/methodology.
Every community benchmark that lands here replaces a bandwidth- derived estimate with a measured number across the site's flagships — same model × hardware × quant combo gets its confidence chip upgraded from E → M.
Measured tok/s on each hardware card replaces the extrapolated number
The placeholder row turns into a reproduced measured row
Real tok/s drives the per-million-tokens local cost
Stream rate switches from estimate to measured
Break-even months stop being approximations
Speed row on this hardware switches to measured
Score confidence chip upgrades from E → M for that row
Top-pick model recommendation switches to the actually-fastest
0 public submissions · newest first
| When | Hardware | Model | Tok/s | Operator | Status | ||
|---|---|---|---|---|---|---|---|
The feed is waiting for its first row. Editorial benchmarks exist on the individual hardware pages, but community submissions through this feed haven't landed yet. Your run could be number one. Submit a benchmark → | |||||||
Editorial can approve a submission as plausible, but it stays out of this numeric feed until reproduction evidence lands.
We re-ran the same protocol on equivalent hardware and got within tolerance. Trustworthy.
Two or more independent operators filed matching numbers, and editorial confirmed. Highest confidence tier.