Measured on real laptops.

Every FastFlowLM release is validated on Strix Point and Halo reference designs. We publish the results in docs/benchmarks so teams can compare apples-to-apples.

The runtime extends AMD’s native 2K context limit to 256K tokens for long-context LLMs and VLMs and, in power efficiency tests, consumes 67.2× less power than the integrated GPU and 222.9× less power than the CPU on the same chip while holding higher throughput.