Muse Glimmer puts a useful agent model within laptop memory

A local model becomes much more useful when it leaves enough memory for the rest of your computer.

Muse Glimmer, released August 10, is Meta’s 30-billion-parameter attempt at that balance. Its weights use Apache 2.0 rather than the bespoke licensing attached to earlier Llama releases. The target is an agent that can inspect images, call tools and recover from failures on consumer hardware.

Meta describes a training process that transfers capabilities from Muse Spark, then adds longer-context agent data and further fine-tuning and reinforcement learning. That is a claim about the training recipe; it does not establish that a local agent can reliably complete arbitrary work.

The memory engineering is concrete. Meta says roughly four-bit quantization brings the language-model weights below 20 GB, leaving room for working memory, image processing and a small drafting model within a 24–32 GB configuration. The drafter proposes token blocks that the main model verifies, accelerating generation.

Simon ran an 18.16 GB LM Studio build. His tests included asking a coding agent to investigate authentication in a fresh Datasette checkout and asking the vision model to describe a photograph. He published the transcripts, making the behavior inspectable. The familiar cycling-pelican drawing was much less convincing: recognizable pieces, awkwardly assembled.

These are useful probes, not a comprehensive evaluation. They test different things: navigating a real repository, understanding a picture and constructing an SVG scene. A model can perform differently across all three.

For a personal agent, local execution also changes the deployment choice. The model can operate without sending every inference to a cloud service, although any tools it calls still need their own permissions and network boundaries. The practical question is which jobs this particular model can complete reliably on the hardware already available.