Qwen3.8 125B (6B active) Flash-Next

User reports Qwen3.8 Flash-Next at 70.0 t/s generation and 1,386 t/s prompt processing on 4 Radeon Pro V620 GPUs at 32K context on the coding suite. Setup is vLLM with the Intel/Qwen3.8-Flash-Next-W4A16-AutoRound quant, using a custom RDNA vLLM fork. At 64K context the coding suite reached 72.5 t/s generation and 1,369 t/s prompt processing; at 128K it reached 68.3 t/s and 1,296 t/s. The regular suite measured 57.9, 59.5, and 56.1 t/s generation at 32K, 64K, and 128K respectively.