Lemonade Server2026.45.3

Lemonade Server runs AI inference on your own hardware, providing local chat, coding, speech, and image-generation services instead of sending those workloads to a remote service. The server also supports multimodal workloads including large language models, speech-to-text, text-to-speech, audio generation, image generation, and 3D generation.

It includes more than 77 built-in models in GGUF, FLM, and ONNX formats, managed with the lemonade-server pull and lemonade-server list commands. Hardware detection selects among CUDA, ROCm, Vulkan, NPU, and CPU backends, with the required backends downloaded on demand. Local models can be combined with cloud offload to OpenAI-compatible providers, while MCP client and server support connects model workflows to Model Context Protocol tools. A Prometheus metrics endpoint supports monitoring, and the background service restarts itself after failures.

AMD acceleration covers RDNA2, RDNA3, and RDNA4 GPUs, RX 6000/7000/9000 cards, Strix Point and Strix Halo APUs, and Instinct MI100, MI200, MI300X, and MI350X hardware through ROCm; AMD XDNA2 Ryzen AI NPUs handle language-model and speech inference. NVIDIA Turing and newer GPUs use CUDA, cross-vendor GPUs can use Vulkan—including Intel and Qualcomm Adreno on ARM64—and systems without GPU acceleration can fall back to the CPU.