An AI model can run on PCs, helping users handle tasks, remember information, use tools and work through multiple steps.
Meta Superintelligence Labs has released Muse Glimmer, a 30-billion-parameter dense open-weights model designed for agentic AI workloads. The model is available under the Apache 2.0 licence and can run locally on AMD Ryzen AI Max+ processor-based systems and workstations equipped with a single AMD Radeon AI PRO R9700 graphics card.
The local deployment is aimed at agentic applications that need long context, persistent memory, tool use and multiple-step workflows. Running these workloads locally can also reduce dependence on cloud services, avoid sending data to remote servers and eliminate recurring token costs.
AMD says Muse Glimmer 30B can run at up to 24 tokens per second on an AMD Ryzen AI Max+ 395 processor. On a system with a single Radeon AI PRO R9700 graphics card, the model reaches up to 53 tokens per second when dFlash speculative decoding is enabled.
The preliminary results were measured on Windows using llama.cpp with the Vulkan backend and dFlash. AMD says further software and model optimisations are expected to improve performance.
Muse Glimmer is designed for tasks that require an AI agent to maintain context, make decisions, call tools, inspect results and continue working across multiple steps. It can also manage memory, recover from failures and continue workflows across restarts.
The model is also designed with local agentic use cases in mind, where AI systems may access files, messages, credentials and other user data. Meta says Muse Glimmer is trained to reduce oversharing, resist prompt injection from untrusted content and respect information boundaries. Local execution gives developers more control over where data is processed.
The Apache 2.0 licence allows developers to use, modify and redistribute the model commercially, making it suitable for integration with existing open-source tools and applications.
For users who want to try the model without building a local inference setup, LM Studio provides a direct way to download and run Muse Glimmer 30B. AMD recommends systems with more than 32GB of VRAM or Variable Graphics Memory. Ryzen AI Max+ processor-based systems and Radeon AI PRO R9700 graphics cards with 32GB of memory are among the recommended configurations.
