Noesion Radarnoesion.ai
Release2026-08-10

Meta released Muse Glimmer, a 30-billion-parameter model built for agent tasks that runs on a single consumer graphics card or a recent Mac, and published the weights under a permissive Apache 2.0 licence. It handles tool calls, multi-step planning and images, and was compressed to roughly a quarter of its full memory size to fit on such hardware. Optimised integrations for common local runtimes are still days away.

What changed

Meta's agentic models required cloud infrastructure and a network connection, and its open releases did not target local agent workflows.

What it unlocks

Running a tool-calling, image-reading agent entirely on a single consumer GPU or a recent Mac, with no internet connection.

  • 30 billion parameters
  • under 20 GB after 4-bit compression
  • up to 3.1x faster decoding
  • Apache 2.0 license

What you need to act on it

  • a machine with roughly 24-32 GB of GPU or unified memory
  • local runtime such as llama.cpp, MLX, Ollama or LM Studio (some integrations arriving in coming days)

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.