For those dabbling with generative AI on the weekends, the AMD-led Lemonade SDK 11.6 is out today with its newest feature release of this open-source software for running local AI apps with optimized LLMs across CPUs, GPUs, and NPUs.

Lemonade 11.6 adds Muse Glimmer 30B GGUF model support into Lemonade's Llama.cpp catalog with draft decoding, vision, and tool calling support. Mesa AI researchers this week published Muse Glimmer as a 30B open agentic model for local agent workflows. AMD was ready with launch-day support and now Lemonade 11.6 is out with catalog integration for easily making use of this interesting LLM.

Also new to Lemonade 11.6 is a new experimental TheNoise ROCm image-generation back-end bringing Anima and Krea-2 support to AMD Strix Halo and Strix Point iGPUs.

Lemonade's Llama.cpp ROCm back-end is now shipping with support for a number of older Instinct GPUs. The AMD Instinct MI100, MI200, MI210, and MI250 GPUs are now working out-of-the-box with the new Lemonade release on Linux.

Lemonade also brings a alias sub-command for aliasing LLM models to use your own desired names for them, the built-in Lemonade benchmark (bench) can now have opt-in vision benchmark scenarios for multi-modal models, and other improvements.

Downloads and more details on the Lemonade SDK 11.6 release via GitHub .