Meta is back in the open weights game. The company's new model, Muse Glimmer, arrives at 30 billion parameters and — perhaps more significantly — under a clean Apache 2.0 licence. That is a clear step up from the idiosyncratic, much-disputed terms attached to the old Llama models.

Why the licence matters

The Llama series was called "open" for years, but its licence did not meet the actual definition of open source: it carried user-count thresholds, branding conditions and usage restrictions. Apache 2.0 is standard, widespread and close to unrestricted. A company's legal team does not need to run a separate assessment before approving it.

On its own, that may do more for adoption in commercial settings than any benchmark number.

What it is optimised for

Meta highlights three things:

  • End-to-end agentic task completion — success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench, which measure the ability to work within scaffolds, write and debug code, and resolve multi-turn requests from start to finish.
  • Reliable tool use — invoking functions with precise schemas throughout extended workflows.
  • Multi-step reasoning — sustaining coherent plans over long horizons.

The ordering is itself a statement of intent: Meta is positioning this model on agentic work rather than conversational quality.

Size is the local-model story

Simon Willison, who tried the model, reports using the roughly 18.2 GB build from LM Studio. The practical meaning of that size: on a machine with 32 GB of RAM or more, the model leaves plenty of space to keep other applications running alongside it.

The friction in local model use is often not raw capability but this balance. A model that swallows a machine's entire memory tends not to get used day to day, even when it technically runs.

Willison also pointed his coding-agent plugin at a fresh checkout of Datasette with the prompt "how does auth work?", and the model produced an answer at the end of a long transcript of tool calls exploring the codebase.

Vision is included

Glimmer is a vision model as well. Asked to describe a nature photograph, it produced a detailed account of a rocky breakwater shoreline with two brown pelicans, identified down to the species name.

Taken together: 30B size, an open licence, an agentic focus and multimodal capability in one package. Meta's return to open weights makes this one of the more concrete options in months for developers running models locally.