Qwen3.6-35B-A3B streams fully on the Ryzen AI NPU

The published numbers

AMD's official FastFlowLM run lists Qwen3.6-35B-A3B at 13.65 tok/s decode and 221.96 tok/s prefill, with the model running fully on the NPU, on Kraken Point. Before we treated any of it as real we checked the run against the official Qwen config, and the numbers line up with a model of that shape.

Decode versus prefill

The two figures are not two speeds for the same job. Prefill ingests the prompt in bulk, many tokens in one pass. Decode emits the answer one token at a time. On the same silicon those are different workloads, and quoting them side by side without saying which is which hides more than it reveals.

Metric	Value	What it measures
prefill	221.96 tok/s	prompt ingestion, many tokens at once
decode	13.65 tok/s	steady-state generation, one token at a time
first pass, ours	75.65 tok/s	llama.cpp Vulkan on Strix Halo

Decode is the number you feel

For streaming output, decode is the number that matters, and 13.65 tok/s is a working stream. The point of the exercise is not that the NPU outruns a discrete GPU, it is that a 35B model runs end to end on the low-power unit that ships inside the machine, with no GPU in the path.

Our first pass

For the same model we measured 75.65 tok/s as a first pass through llama.cpp Vulkan on a Strix Halo. It is a first pass, not a comparison: different backend, different silicon, different kernel story. The honest reading is that both paths move this model at usable speeds, and that each side is early and improvable.

The record

AMD's numbers decode cleanly, the model checks out against the official Qwen config, and our own first pass lands in the same usable range. The harness and the config checks are in the repo, and we will be the first to correct this entry when a newer pass beats it.