Decode · tok/s
100
Prefill
—
tok/100W
28.6
Really nice work @coldniko! Here's my little contribution for Qwen3.8-Flash-Next (MoE, IQ3_S) on ONE RTX 3090 with the Strata engine (+9 à +14 % gains): ▪️ 61 → 100 tok/s on a short question ▪️ 78 → 99 tok/s with a 56K-token prompt ▪️ 62 → 82 tok/s at 256K context (98% full)
First-person single 3090. Rank is the quoted 100 on a short question after the patches. 99 and 82 stay in the note. Prior GSQ-RCO pair on the same chip was 72 at 128K.
