AMD has pushed out Lemonade 11.5. The local AI server update arrived just one week after version 11.0. It landed during the company’s Advancing AI event. The release marks a clear step forward for developers who want fast, private model inference without cloud bills or data leaks.
The project, backed by AMD and hosted openly on GitHub, lets users run large language models across Ryzen AI processors, Radeon graphics cards and standard CPUs. It works on Windows and Linux alike. Phoronix first reported the news. The story highlighted how the software simplifies deployment of demanding workloads that once required specialized setups.
At its heart Lemonade acts as a unified server. It speaks the OpenAI API language. Applications built for ChatGPT or similar services can point at a local endpoint instead. No code changes needed. Earlier AMD technical articles explained the approach. The company positioned Lemonade as a way to bring neural processing units, graphics processors and central processors together in hybrid mode for better speed and lower power draw. One piece from AMD noted that the server offers “the only open-source OpenAI-compatible server to offer Ryzen AI NPU acceleration of LLMs” at the time of its initial launch.
Yet the real story in 11.5 sits with the newly completed Lemonade Router. Name a model collection.router and incoming requests get steered automatically. Rules. Classifiers. Semantic similarity. Even an LLM-as-router option where a small model decides the best target. Each decision leaves traces so engineers can debug why a query went one way instead of another. A graphical interface now lets users author and edit these routing policies without touching JSON by hand. A new POST /v1/classify endpoint handles text classification. It runs on a fresh ONNX Runtime backend.
The router turns a simple server into something closer to an intelligent traffic cop for AI workloads.
That matters. Enterprises testing multiple specialist models suddenly gain an easy way to route customer-support queries to one model, code questions to another and creative tasks to a third. No custom middleware required. The official release notes on the Lemonade GitHub repository put it plainly: “The Lemonade Router is complete: name a collection.router model and requests are steered by rule, classifier, semantic-similarity, or LLM-as-router policies, with per-candidate decision traces, a GUI for authoring and editing routers, and a POST /v1/classify text-classification endpoint backed by a new onnxruntime backend.”
Version 11.5 adds more than routing. A server-side job engine accepts multi-step recipes from clients. Those recipes live at dedicated /jobs endpoints. Operators can pause them, interrupt them, resume or delete them on demand. The feature opens the door to long-running agentic workflows that survive browser refreshes or network hiccups. Meanwhile the lemond daemon can now host MCP clients. It connects to external stdio-based MCP servers and exposes their tools through gated internal endpoints. Security teams will notice the breaking change too. The server no longer sends a wildcard Access-Control-Allow-Origin header by default. Remote browser clients must be listed explicitly in an environment variable or they receive a 403.
Installation stayed straightforward. Windows users grab a signed .msi package. Linux distributions from Ubuntu to Fedora offer native packages or Docker images. macOS gets a .pkg installer. Embeddable builds target developers who want to bundle the server inside larger applications. The project credits a long list of contributors. Several names from AMD appear among router milestone authors, including eddierichter-amd. Community members supplied fixes for everything from Arch Linux paths to macOS unit tests.
But why does this matter now? Local AI has moved from hobbyist experiment to enterprise requirement. Data-privacy rules, latency demands and ballooning cloud costs push companies toward on-premise inference. AMD sees the opening. Its Ryzen AI 300-series chips ship with dedicated neural processing units. Radeon discrete cards bring massive parallel compute. Lemonade ties those assets together so developers don’t have to choose one accelerator and rewrite code when hardware changes.
An AMD article from earlier this year spelled out the vision. It described Lemonade as “a unified API for local AI developers” built in lightweight C++ that installs quickly across operating systems. The stack auto-detects available hardware and picks the best backend. llama.cpp for broad compatibility. Ryzen AI libraries for NPU acceleration. ROCm paths for high-end GPUs. The result feels like a single coherent service even when the underlying silicon shifts.
Recent coverage echoes the momentum. A July 2026 post on daily.dev discussed the 11.0 release and its text-to-speech additions. It framed the project as AMD’s answer to running capable models on Ryzen CPUs, Radeon GPUs and the latest NPUs under both Windows and Linux. Community chatter on X has picked up the thread too. One developer noted that Lemonade combined with FastFlowLM now enables 256k context windows on Ryzen AI silicon without a discrete card. The Linux NPU driver work that took two years finally pays off.
Still, gaps remain. The 11.5 notes carry no new benchmark numbers. Earlier AMD documents showed impressive results on specific models such as Gemma-3-4B for design assistance or Qwen3-Coder-30B for programming tasks. Those figures depended on exact hardware pairings. Real-world performance will vary with model size, context length and whether the workload hits NPU, GPU or falls back to CPU. Power efficiency numbers also stay scarce in public materials. That leaves system integrators to run their own tests.
Security updates in this release deserve attention. Contributor-hosted backend assets now use pinned SHA-256 checksums. HTTP trust policies apply to POST paths. These changes close potential supply-chain vectors that have bitten other open-source AI projects. The decision to tighten CORS rules reflects growing production use. What began as a developer convenience now runs in environments where cross-origin risks cannot be ignored.
Look ahead. The router’s classifier and semantic-similarity options open new application patterns. A single endpoint could serve an entire AI suite. One model answers simple questions. Another handles complex analysis. A third generates images. The server decides on the fly. Add the job engine and developers can chain these steps into persistent agents that monitor, reason and act without constant human supervision. The MCP client support hints at deeper tool-calling integration. External services become discoverable tools inside the Lemonade environment.
AMD’s investment looks strategic. The company faces stiff competition from NVIDIA’s software stack and Intel’s own AI PC push. By funding an open project that runs well on its silicon, AMD attracts developers who might otherwise default to CUDA. The community contributions listed in the release show the bet is working. New faces appear alongside AMD engineers. The project moves faster than a purely internal effort could.
Critics might argue that OpenAI API compatibility is table stakes. Many local servers offer it. What sets Lemonade apart is the hardware-specific tuning and the emerging routing intelligence. Few competitors expose NPU acceleration so cleanly on AMD platforms. Fewer still ship a router with decision traces and a GUI. Those details could decide who wins the local AI server race over the next year.
Downloads are live. The GitHub release page lists every package. Documentation walks through router configuration and job API usage. For teams already running 11.0 the upgrade path looks painless. For newcomers the one-click Windows installer removes most friction.
The pace surprises. Two major releases in two weeks. Each adds production-grade features. AMD and its open-source partners show they treat local AI as a priority, not a side project. The next versions will likely expand model support, add more backends and tighten performance on the latest Ryzen and Radeon hardware. For now 11.5 gives developers a smarter, more controllable foundation for private AI deployments. That alone makes it worth a close look.
WebProNews is an iEntry Publication