It's really not shabby at all:
Qwen3.8-Flash-Next 125B MoE NVFP4 (nvidia/Qwen3.8-Flash-Next-FP4-Local) on LIL blackwell-llm-docker (release-10152026), container qwen38-fp4, TP2 across both GPUs (~94GB each), 262K ctx, MTP x3, fp8 KV, PLE CPU-offload. OpenAI-compatible API :8100. Start/stop via wrapper unit qwen38-fp4.service (docker start).