Status

Reproduced from third-party benchmark publishers with attribution, never relicensed. None of it was measured by us, and we say so wherever it appears. It exists so that our own modeled figures can eventually be scored against something real rather than merely banded.

No first-party measured performance

We now reproduce third-party measured benchmarks with attribution — image generation throughput per GPU, and speech recognition accuracy and speed. But every figure WE compute is still modeled from memory bandwidth with a published error band. Nothing on this site has been measured by us.

Measured data is not redistributable

The sources we reproduce publish no licence of their own. We display and attribute them, and exclude them from our CC-BY bulk dump, which therefore contains only data we derived ourselves.

Image throughput is community-submitted, not controlled

The image-generation figures per GPU aggregate thousands of community runs across different models, resolutions and step counts. Read the spread, not the median alone — it is not one controlled configuration.

Video has no throughput data at all

Component sizes are exact. Seconds per clip on consumer hardware is unpublished anywhere credible, so we show nothing rather than a guess.

The catalog is a slice, not the whole Hub

We index the most-downloaded base models. The full deduplicated set is roughly 1,500–3,000.

Gated models may lack architecture

Exact file sizes are always available, even for gated repositories. Architecture sometimes has to come from an ungated mirror, which we flag with reduced confidence, or is unavailable entirely.

Compute buffer is modeled with constants we chose

There is no closed form for the runtime's working buffers. Ours scales with the terms that should drive it — feed-forward width, micro-batch, and context without flash attention — but its constants are not fitted against measured allocations. It is the widest error band on the site and the reason a total carries one when its parts do not.