A user tested the 125B parameter Qwen3-8B-Flash-Next MoE model on a 128GB Ryzen AI Max+ 395 system, finding it consumes 103.7 GiB VRAM. Decode speeds reached 34.1 tok/s (GPU only), 53.7 tok/s with MTP, and 34.2 tok/s using NPU offload, suggesting diminishing returns for massive local models.