FastFlowLM runs Qwen3.5-4B on AMD NPU

FastFlowLM is making AMD NPU laptops feel like a real local-AI tier, not a novelty. The interesting part is the mix of low thermals, tool-calling, and multimodal support, which is what turns a demo into something developers can actually build on.

• –15 tok/s decode at 1k context is respectable for a laptop NPU, but the 9.6 tok/s figure at 32k shows long-context overhead still matters.
• –Prefill lands at 378-493 tok/s, and image TTFT is 3.7s at 720p versus 7.5s at 1080p, so lightweight vision workflows are viable.
• –Tool-calling support makes this relevant for agents and local copilots, not just benchmark bragging rights.
• –Support across all XDNA 2 NPUs makes this a platform story, and the 256k-token ceiling matters more on bigger-memory systems than this 32GB laptop.