Versions:
Lemonade Server, developed by AMD and currently at version 10.0.0, is a server interface that exposes local large-language-model inference through the same OpenAI-compatible REST endpoints used by cloud services, enabling chat clients, coding assistants, research tools, and other AI-powered applications to switch from remote endpoints to private on-device execution without code changes. By offloading compute to the PC’s integrated or discrete GPU and to the dedicated Neural Processing Unit found in recent Ryzen AI processors, the software delivers token-generation speeds that routinely exceed 30–40 tokens per second for 7–8-billion-parameter models while keeping all prompts, context windows, and fine-tuned weights entirely within the user’s machine. The package ships in fifteen incremental releases, each expanding the roster of supported quantized formats—GGUF, ONNX, and DirectML among them—and adding runtime optimizations such as speculative decoding, dynamic batching, and automatic KV-cache compression that cutVRAM use by up to 35 %. Typical deployment scenarios include privacy-sensitive enterprises that prohibit external API calls, developers prototyping offline copilots, students running Llama- or Mistral-based tutors on laptops, and content creators who need a locally hosted brainstorming partner that functions even without Internet connectivity. Configuration is handled through a single JSON file where model paths, context lengths, and concurrency limits are declared; once the service is started, any application that speaks the familiar /v1/completions or /v1/chat/completions interface can direct traffic to localhost:8080 and receive streaming JSON responses indistinguishable from those of commercial endpoints. Lemonade Server is available for free on wget.nero.com, with downloads provided via trusted Windows package sources (e.g. winget), always delivering the latest version, and supporting batch installation of multiple applications.
Tags: