AMD Lemonade 11.5: Local AI Server Equipped with Router Functionality
AMD-backed open-source local AI server "Lemonade" updated to v11.5. The new version introduces Lemonade Router, multi-step jobs, and MCP client functionality.
AMD announced the release of the latest version of its open-source local AI server, “Lemonade” v11.5, on July 22, 2026, during the AMD Advancing AI event. This update comes just a week after the release of version 11.0. Lemonade has garnered attention as software that allows for easy deployment of large language model (LLM) workloads on Ryzen AI NPUs, Radeon GPUs, and CPUs, supporting both Windows and Linux environments.
The Major Update: Completion of the Router
The most significant update in Lemonade 11.5 is the completion of the Lemonade Router. This router features the ability to automatically route queries to the most suitable model based on predefined policies. Four routing methods are available: rule-based, classifier-based, semantic similarity-based, and “LLM-as-a-Router,” which uses an LLM itself as the router.
In the LLM-as-a-Router model, a smaller LLM analyzes the query content and determines the most appropriate model to handle it. For instance, image generation queries could be routed to Stable Diffusion, while text summarization queries could go to Llama. Such collaborative operations across multiple models can be achieved simply by defining policies.
Policies can be defined either through JSON files or directly via Lemonade’s GUI, enabling even non-engineer users to intuitively set up routing configurations.
Introduction of a Server-Side Job Engine
Lemonade 11.5 also introduces a new server-side job engine. This feature allows clients to submit multi-step workflows (recipes consisting of multiple procedural steps) to the server and manage them via the /jobs endpoint.
Specifically, users can pause, interrupt, resume, or delete multi-step recipes through the API. This means that complex workflows, such as those involving multiple stages of inference pipelines using different models, can now be fully controlled on the server side.
MCP Client Host Functionality
Another new feature is the ability for the Lemonade daemon (“lemond”) to act as an MCP (Model Context Protocol) client host. This enables connections to external stdio-based MCP servers and allows Lemonade’s agent flow to utilize the tools and data sources provided by these servers.
MCP, a standard protocol proposed by Anthropic, connects AI models with external tools and data sources. With Lemonade now equipped with MCP client host functionality, local AI servers can access external features such as file system operations, database queries, and web API calls directly from models.
Maturity as a Local AI Execution Platform
According to a report by Phoronix’s Michael Larabel, Lemonade is a central component of AMD’s local AI ecosystem. It transparently utilizes NPU, GPU, and CPU hardware resources, allowing users to run LLMs without being concerned about hardware differences.
With the completion of the router in v11.5, Lemonade has evolved from a single-model execution server into a policy-based AI workflow execution platform capable of orchestrating multiple models. The combination of multi-step jobs and MCP compatibility demonstrates the potential for achieving end-to-end automation in local environments.
Editorial Opinion
The release of Lemonade 11.5 marks a significant milestone in the evolution of local AI execution platforms. With the completion of the router, edge environments can now achieve model selection and routing flexibility on par with the cloud.
In the short term, this tool has the potential to dramatically improve prototyping speed for developers utilizing AMD’s Ryzen AI-equipped PCs or Radeon GPUs. The LLM-as-a-Router feature, in particular, offers efficiency in both power consumption and latency by leveraging smaller models to categorize queries, eliminating the need for large models to run constantly.
From a long-term perspective, the implementation of MCP client host functionality positions Lemonade to evolve from a mere inference server into a control platform for local AI agents. If external tool integration becomes standardized, developers could create self-contained agent systems entirely within local environments. Whether AMD will strengthen this open-source strategy to differentiate itself from competitors like Intel and NVIDIA could significantly impact its position in the market.
References
- “Lemonade 11.5 Local AI Server Released With Completed Lemonade Router”, by Michael Larabel — Phoronix, 2026-07-22T22:35:35.000Z (ARR)
- Source URL: https://www.phoronix.com/news/AMD-Lemonade-11.5
Frequently Asked Questions
- What skills are required to use Lemonade 11.5's router functionality?
- Basic usage is possible if you can define policies in JSON. Settings can also be configured through the GUI. While a small LLM is required for the LLM-as-a-Router feature, programming is not strictly necessary to use the router functionality itself.
- What specific use cases are envisioned for the multi-step job engine?
- For example, tasks such as recognizing text via OCR, translating it into English using a translation model, and summarizing it with a summarization model can be registered as a single recipe on the server and executed or managed via the API. It is particularly suited for automating inference pipelines.
- What are the benefits of connecting to external MCP servers with the MCP client host functionality?
- It enables AI models to directly perform tasks like file operations, database queries, and external API calls. For instance, an agent model running on Lemonade can perform tasks while reading and writing local files via MCP.
Comments