AMD has introduced a localized computing platform designed specifically for the Agentic PC. The ecosystem combines high power Ryzen AI processors and Radeon graphics cards with Meta's new Muse Glimmer 30B open weights model. But running these massive models locally requires hardware with substantial memory capacity.
Meta Superintelligence Labs recently released Muse Glimmer 30B under the Apache 2.0 license. This dense open weights model targets workflows requiring long term memory and multi step execution. Unlike basic chat programs, the system manages context over extended sessions and recovers from errors automatically. Developers can modify and distribute the model freely for commercial software builds.
Running agentic tasks locally keeps sensitive files, messages, and login details private. AMD hardware manages this processing load without the lag or recurring costs of cloud systems. Recent benchmark tests conducted by AMD show the model reaching up to 24 tokens per second on a Ryzen AI Max+ 395 processor with 128GB of memory. Stepping up to a dedicated Radeon AI PRO R9700 graphics card with 32GB of VRAM increases performance to 53 tokens per second. These initial trials used the popular llama cpp framework on Windows 11 Pro with speculative decoding active.
 |
| Meta's Muse Glimmer 30B running on an AMD Ryzen AI Max+ in llama.cpp |
Getting the model running takes only a few minutes. Consumers can use LM Studio to discover, download, and run Muse Glimmer on supported AMD computers. Systems need more than 32GB of memory to run the model smoothly. Power users can run a local server and route the connection to other independent agents like Hermes Agent or Open Claw. This allows for immediate testing against active office files and software tools.
Integrating this capability into commercial software requires different tools. A program called Lemonade packages the model into a tiny 4 MB binary. This allows developers to embed the entire AI system directly into their own applications. The software acts as a private local service, running optimized inference across Ryzen AI processors and Radeon graphics cards. End users do not need to install separate frameworks or configure complex settings to run the software. It changes the model from a standalone test into a core application asset.
