Qwen3.8 27B hits 140 t/s on AMD RX 7900 XTX with speculative decoding

Jebbyk1 · reddit · 2026-11-20

A Reddit user ran Qwen3.8 27B on an AMD RX 7900 XTX at 100-140 tokens/s using a llama.cpp fork tuned for the RDNA3 architecture with speculative decoding.

Per the llamaperf community report: Q4KM quantization, 192K context, 140.0 t/s generation and 862.3 t/s prefill — showing RDNA3 GPUs are now quite viable for local LLM inference.

Original post →

More from Infra

Infra channel →