On August 10, Meta Superintelligence Labs put Muse Glimmer on Hugging Face – under an Apache 2.0 license, so it’s free to use, commercial work included. The interesting part isn’t the size. It’s where the model fits: onto a single graphics card you can keep at home.
The memory trick
Muse Glimmer has 30 billion parameters. At full precision that would need around 55 GB of memory – well beyond what sits in a normal PC. Through 4-bit quantization, Meta pushes the requirement down to 18–20 GB. That lets the model run in a 24 or 32 GB VRAM budget, on a consumer GPU or a Mac. Offline. Without a single token leaving your machine.
Glimmer is distilled from the larger teacher model, Muse Spark. Meta pitches it for local coding, function calling, and autonomous agent workflows: reliable tool use, multi-step reasoning, failure recovery, and multimodal input. It’s also built to plug into common agent orchestrators like OpenClaw.
Why this is more than another release
The market for coding agents currently runs almost entirely through the cloud – Claude Code, Codex, Meta’s own Muse Code. You send your code to a provider, the provider sends answers back, and the bill grows with every task. A model that works agentically on a local 24 GB card flips that logic.
For anyone with sensitive code, compliance worries, or simply no desire to pay per token, that matters. Meta is positioning itself squarely against the closed models from OpenAI and Anthropic – not on peak performance, but on access and control.
My take
Whether Glimmer keeps up with the cloud models is something daily use will have to prove. 30B on a desk isn’t 400B in a data center, and 4-bit quantization costs precision. But maybe that’s not the point.
The interesting question is a different one: how much agent work can you get done on hardware that’s already on your desk? If the answer is “more than you’d think,” it shifts the math for a lot of teams. And Meta is shipping the building blocks for it in the open.
Sources: