AMD Launches Lemonade, an Open-Source Local LLM Server for GPUs and NPUs
AMD has released Lemonade, an open-source server for running large language models locally on consumer hardware, according to the project's official site, lemonade-server.ai. The tool is designed to let developers run LLMs on-device by splitting workloads between a machine's GPU and its NPU (neural processing unit), rather than relying solely on one type of processor.
The NPU angle targets AMD's Ryzen AI chips, which include on-chip neural accelerators built for efficient, low-power inference. According to Lemonade's documentation, the server can delegate work between GPU and NPU to balance raw performance against power draw — a combination aimed at laptops and other AI PCs where battery life and thermals matter as much as speed.
Lemonade exposes models through an API-compatible interface, so developers can plug local inference into existing applications without depending on cloud services from providers like OpenAI or Anthropic. Running models locally avoids per-request API costs and keeps data on-device, though it requires compatible AMD hardware to take advantage of the NPU offload.
The project's launch was discussed on Hacker News, drawing developer interest in AMD's push to build tooling around its Ryzen AI hardware and ROCm software stack — an area where Nvidia's CUDA ecosystem has held a long-standing advantage. Documentation and installation instructions are available on the project's website.