Decode · tok/s
42.1
Prefill
—
tok/100W
7.3
I made some improvements on Qwen3.8-Flash-Next on crow-nest (RTX 5090, 64 GB DDR5) Decode, before → after: Average: 37.9 → 42.1 tok/s (+11.1%) Median: 37.4 → 41.4 tok/s (+10.6%)
First-person RTX 5090 + 64GB DDR5. Rank is the quoted average 42.1 after improvements. Quant not named. Runtime is crow-nest.