AMD Lemonade 11.6 adds Muse Glimmer 30B and experimental ROCm image generation

AMD has released Lemonade SDK 11.6, the latest update to its open-source toolkit for running local AI applications with optimized large language models across CPUs, GPUs, and NPUs.

The headline feature is support for the Muse Glimmer 30B model in GGUF format, now integrated into Lemonade's Llama.cpp catalog. Muse Glimmer is an open agentic model published this week by Mesa AI researchers, designed specifically for local agent workflows. Lemonade 11.6 ships with launch-day support, bringing draft decoding, vision, and tool-calling capabilities to the model. AMD moved quickly to ensure the model was usable from day one, and the new catalog integration makes it straightforward to put this LLM to work.

Another significant addition is an experimental TheNoise ROCm image-generation backend. It enables Anima and Krea-2 model support on AMD's Strix Halo and Strix Point integrated GPUs, opening up local image synthesis on these platforms.

The Llama.cpp ROCm backend has also been expanded to cover a range of older Instinct accelerators. The AMD Instinct MI100, MI200, MI210, and MI250 GPUs now work out-of-the-box with Lemonade 11.6 on Linux, which should appeal to users of previous-generation data center hardware.

Beyond these major items, the release includes several quality-of-life improvements. A new alias sub-command lets users assign custom names to LLM models, making them easier to manage in daily workflows. The built-in Lemonade benchmark tool, bench, now offers opt-in vision benchmark scenarios for multi-modal models, allowing users to evaluate vision-language capabilities alongside traditional text performance.

Lemonade SDK 11.6 is available for download on GitHub, where the full changelog and installation instructions are also provided. The project continues to position itself as a flexible, vendor-backed option for developers and enthusiasts who want to run AI models locally without relying on cloud services.