Meta Superintelligence Labs has released the 30-billion-parameter agentic AI model Muse Glimmer, made available under the permissive Apache 2.0 open-source licence, that runs locally on consumer desktops and laptops without relying on cloud infrastructure.
Muse Glimmer has been optimised for always-on local agent workflows, including function calling, coding assistance, multi-step reasoning, and multimodal understanding. At 4-bit quantisation, the model fits comfortably within 24GB to 32GB of video memory, making it suitable to run a single consumer-grade GPU, whether on Windows or macOS, not necessarily an AI PC.
Built to work with the tools developers are using, Muse Glimmer lets users run AI anywhere, anytime, with or without an internet connection.
The model is being released on Hugging Face, with optimised builds for llama.cpp, MLX and ExecuTorch due in the coming days. Partners such as Ollama, LM Studio, Unsloth, vLLM, and SGLang are supporting the release, while AMD has also announced native optimisation across its latest hardware.
What sets Muse Glimmer apart from similarly sized models is its agentic design, which give it the ability to sustain multi-step workflows, recover from errors and chain tool calls over extended sessions.
It supports interleaved text-and-image input via a dedicated perception encoder, making it capable of interpreting screenshots, charts, and documents alongside conversation. Trained on data from more than 100 languages, it also offers broad multilingual capability.
The release marks a notable step toward local, persistent AI agents that stay on the device rather than in the cloud. With privacy-sensitive data never leaving the user’s machine, and no ongoing API costs, Muse Glimmer could accelerate adoption of agentic AI in enterprise, developer and consumer environments alike. It could also further boost the demand for AI PCs.
AMD hardware support
AMD has moved quickly to validate Muse Glimmer on its platforms, reporting preliminary performance figures that suggest strong real-world usability. On an AMD Ryzen AI Max+ 395 processor, the model delivered up to 24 tokens per second, while a single Radeon AI PRO R9700 graphics card pushed throughput to 53 tokens per second with dFlash speculative decoding enabled. Testing was conducted on Windows using llama.cpp with the Vulkan backend.
Recommended hardware includes Ryzen AI Max+ systems or Radeon AI PRO R9700 (32GB) cards, with at least 32GB of graphics memory advised for out-of-the-box operation.
Entelechy Asia 