Speedy is shockingly good for qwen3.8 FN
retained text
I had *just* finished setting up avalon, then i found speedy, which crushed avalon. https://github.com/freriadoc/Avalon https://github.com/DevDock-AI/Speedy Hardware: 4090, 128gb ddr5 3600, ryzen 9 7950x ISTA IQ3_S quant: | Context | Reading: cold prefill s | Reading: reported tok/s | Writing: generation tok/s | Whole exchange: wall s | | ------: | ---------------------------------------- | ------------------------------------ | ----------------------------------------- | ---------------------------------------- | | 60K | Strata 54.7 / Avalon 22.4 / Speedy 13.7 | Strata — / Avalon 2678 / Speedy 4389 | Strata 75.0 / Avalon 102.5 / Speedy 110.0 | Strata 66.6 / Avalon 32.1 / Speedy 23.9 | | 120K | Strata 110.7 / Avalon 42.1 / Speedy 26.0 | Strata — / Avalon 2849 / Speedy 4622 | Strata 71.8 / Avalon 101.6 / Speedy 105.8 | Strata 123.2 / Avalon 52.0 / Speedy 37.0 | | 198K | Strata 187.2 / Avalon 71.2 / Speedy 44.1 | Strata — / Avalon 2779 / Speedy 4494 | Strata 71.9 / Avalon 103.0 / Speedy 104.3 | Strata 199.8 / Avalon 81.5 / Speedy 56.1 | | Context | Writing Strata -> Speedy | Writing Avalon -> Speedy | Reading Strata -> Speedy | Reading Avalon -> Speedy | Wall Strata -> Speedy | Wall Avalon -> Speedy | | ------: | -----------------------: | -----------------------: | -----------------------: | -----------------------: | --------------------: | --------------------: | | 60K | +47% | +7% | 4.0x | 1.6x | 64% shorter | 26% shorter | | 120K | +47% | +4% | 4.3x | 1.6x | 70% shorter | 29% shorter | | 198K | +45% | +1% | 4.2x | 1.6x | 72% shorter | 31% shorter |
document fetched 2026-11-02 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1xlfhrd/speedy_is_shockingly_good_for_qwen38_fn