OpenNPU
Local AI on the NPU of AMD Ryzen AI laptops, on Linux: chat, vision, voice and generation with its own open NPU engines.
Intention
On Linux, the NPU in millions of Ryzen AI laptops mostly sits idle. OpenNPU is the app that runs your own models on it, with open engines and no closed runtime underneath.
What exists today
An open-source app (Apache-2.0, beta v0.3) with a web UI, a GTK desktop app and a terminal chat. It loads GGUF files from LM Studio or Hugging Face and Hugging Face checkpoints, supports vision, voice and image, video and speech generation, keeps a local usage dashboard, and exposes an OpenAI- and Ollama-compatible API.
Technical approach
Own NPU engines compiled per model shape with the MLIR-AIR toolchain. llama.cpp matrix math is offloaded to the NPU. Runs on XRT 2.25 with kernel 6.17.
Evidence and known limitations
- Tested on one machine: a Ryzen AI 7 PRO 350 (XDNA2). Other XDNA2 chips should work but are untested.
- Beta software. See Limits in the README.
- Related benchmarking work (npu-bench, Laya on NPU) is not yet public.
Wider hardware coverage as reports come in.