FastFlowLM 1.0.8 brings MiniCPM vision to AMD NPUs
FastFlowLM released version 1.0.8 on October 12, adding MiniCPM V4.7 1B support and preview HRX packages for Windows and Linux. The update expands the choices for running AI on compatible AMD neural processing units.
The vision model supports optional thinking and tool calling, with a default context of 32,000 tokens that can extend to 256,000. This integration accepts images and text. Video and audio input are unsupported. HRX remains a preview, while XRT stays the default runtime.
The release also includes an optional source build for the Gemma 4 E2B and E4B engine. That code was merged on October 8. Other engines, including the Flash and 12B variants, remain prebuilt.
What the project measured
The project’s version specific benchmarks used a Ryzen AI 5 340 system with 64GB of memory and the default Performance power mode. MiniCPM generated 47.7 tokens per second at a context length of 1,000 tokens. At 32,000 tokens, that fell to 23.8. Prompt processing reached 1,504.7 tokens per second for a 1,000 token input and 1,666.3 at 32,000.
An image description test recorded 2.1 seconds until the first token at 720p and 2.5 seconds at 1080p, 2K and 4K. The page explains that large images hit a fixed tile limit, so higher resolutions add no image tokens. The equal timings therefore do not establish that every extra pixel is being examined.
These are project measurements on one specified configuration. They provide a useful starting point for judging responsiveness, but do not establish answer quality or the results a different computer, prompt or sustained workload will produce.
Hardware limits still apply
The Linux installation guide requires an AMD XDNA2 NPU. Its supported list includes Ryzen AI Max 300, 300 and 400 families, plus Z2 Extreme. Older Ryzen AI 7000, 8000 and 200 series processors use XDNA1 and are excluded. A Ryzen badge alone is therefore insufficient to establish compatibility.
Linux deployment also depends on compatible NPU firmware, the amdxdna driver and adequate memory locking limits. The guide specifies firmware version 1.1.0.0 or later, with kernel 7.0 or newer, or a separate DKMS driver installation. The guide provides a validation command and troubleshooting for cases where the runtime cannot see the device. Those requirements remain relevant before comparing any model speed.
For developers assessing the update, a practical test is to run representative image questions at the context lengths their application actually needs. Record response delay and correctness together. That will show whether the new integration suits the intended workload.
Illustrative laptop photograph by Dan Tabifranca from the WordPress Photo Directory, available under CC0. The photograph predates this release.
