AnalysisDevelopersSeptember 12, 2026

Qwen3.8 Flash Next hits 1.2k t/s prefill on Strix Halo

A working log on Radeon 8060S (gfx1151, ROCm) reports 1.2k t/s prefill for Qwen3.8 Flash Next on Strix Halo. The stack spans a HIP runtime change adding retained PM4 command lists, ROCm work in llama.cpp, and two Qwen3.8 models.

1 source

More stories today

Open the live feed