Qwen3.8-27B on AMD R9700 hits 227 tok/s with lossless block-diffusion drafter

samsja19 · x · 2026-08-27

Lucebox engine now serves Qwen3.8-27B on a single AMD Radeon AI PRO R9700 (32GB, RDNA4) using DFlash2 block-diffusion drafter, achieving up to 227 tok/s on code, 208 tok/s on HumanEval, and 133 tok/s on math. Using Unsloth's UD-IQ4XS quant, it matches 8-bit reference with KL divergence 0.018 and 94% top-1 agreement, and is lossless.

Original post →

More from Infra

Infra channel →