bestmodel.run / cloud anchors

Runs that anchor the scale.

Measured runs on rented GPUs, outside the community hardware pool, to anchor the scale.

sd-turboimagemeasured

6.09 img/s

L4 24GB (modal) · basis measured · n=5

Whisper Large v3audiomeasured

4.82 ×real

L4 24GB (modal) · basis measured · n=5

A10 × L4

Same model, same 4-bit quantization, n=3 on both sides: decode moved from 44.76 to 75.83 tok/s, or 1.69×, while memory bandwidth moved from 300 to 600 GB/s, or 2.0×. The gain is sub-linear relative to bandwidth, consistent with memory-limited decode without being proportional to it.