Meta has released Muse Glimmer, an agent-based model with 30 billion parameters that can run on a single consumer-grade GPU

Thanks to 4-bit quantisation, Meta is able to fit a complete agentic model into 18 to 20 GB of VRAM, making it usable on a standard PC or Mac fitted with a consumer-grade graphics card.

On 10 August, Meta Superintelligence Labs released Muse Glimmer, an open-weight model with 30 billion parameters designed to run AI agents continuously on a single consumer-grade GPU, licensed under Apache 2.0. Thanks to 4-bit quantisation, the model’s memory footprint is reduced from 55 GB to around 18–20 GB, meaning it can fit – along with its cache and speculative decoder – onto a card with 24 or 32 GB of VRAM.

The model is aimed at developers building local agents, code tools or function-calling systems without having to rely constantly on a network connection. It has been trained and evaluated specifically for end-to-end agentic task execution and multi-step reasoning, with weights available on Hugging Face in several formats (BF16, GGUF, ExecuTorch).