Community contribution

Submit your local AI benchmark

Contribute a real measurement from your hardware. After moderation, your submission appears as "Measured on real hardware" on the relevant verdict pages, replacing fit-engine estimates for that model × hardware × runtime combination.

Hardware

Model

Runtime

Workload

Quantization (optional)

Context tokens

Measurements (at least one required)

Decode tok/s

Peak memory (GB)

TTFT (ms)

Your handle (optional)

Notes (optional)

How to measure

We recommend using a standard tool to generate reproducible numbers:

aidatatools/ollama-benchmark — CLI that reports tokens/sec for any Ollama model
rockyRunnr/ollama-bench — includes TTFT, memory peak, JSON export
Ollama's official ollama bench command — CSV output

For Apple Silicon with MLX, mlx-lm reports tok/s directly. Run warm inference (a short warmup prompt first) and report the mid-run decode throughput.

Moderation policy

All submissions go through light manual review before appearing on public pages. We check for: reasonable numeric ranges (not 10,000 tok/s on a laptop), consistent hardware × model memory fit, and no obviously spammy notes. Typical turnaround: 1-3 days. Submissions outside 3σ of community-reported ranges are flagged for additional verification.