Community benchmarks: HartsyInference on real GPUs

HartsyInference runs on whatever GPU the community brings. This page shows the results that have passed review: decode speed, time to first token, peak VRAM, and how far the numbers spread across machines. Every row links to the evidence bundle it came from.

Run the benchmark on your own GPU with the benchmark guide. A result appears here only after it is reviewed, and it refreshes weekly, so a withdrawn result can stay up for up to a week.

Reviewed results from community GPUs, produced with the HartsyInference benchmark runner. Each row is one GPU, the median of its sessions. This page refreshes weekly, so a withdrawn result can stay up for up to a week.

sd15-f16-512