Decode · tok/s
370
Prefill
—
tok/100W
64.3
TensorFold 0.6.6 + DFlash2 Code: 370 tok/s Prose: 174 tok/s Structured output: 634 tok/s
First-person RTX 5090. Rank is the quoted code 370. Prose 174 and structured 634 stay in the note. Quant is MLX format as stated; TensorFold 1.x not yet supporting this model on CUDA.
