Meta released an AI agent that runs on your own laptop
Meta open-sourced Muse Glimmer, a 30-billion-parameter agentic model, under a permissive Apache 2.0 licence on 10 August 2026. Distilled from Meta's closed Muse Spark frontier model, it's built for always-on local agent work. Such as tool calling, local coding, long-horizon reasoning and failure recovery. It is small enough to run on a Mac or PC with a single 24GB consumer GPU using 4-bit quantisation. Weights are on Hugging Face now with Meta reporting it outperforms Gemma4-31B and Qwen3.6-27B in its size class on agentic benchmarks.
What it actually means
For a year the unspoken rule has been that a capable agent needs a data centre behind it. For instance, someone's cloud, someone's per-token bill, someone else's servers holding your users' data. This is the first credible crack in that rule. A model good enough to call tools, write code and recover from its own mistakes now fits on a machine you already own, runs offline, and costs nothing per call once it's downloaded. It won't beat a frontier model on the hard 20%, but the benchmarks are "strong for its size", not best-in-class. It’s great for the always-on, high-volume agent work i.e. watching a folder, drafting, triaging, running local checks. Suddenly the running cost goes to zero and the data never leaves the device. The move this week is to point one real local task at it and watch whether the quality holds; at the same time, keep the cloud model.
A capable agent no longer needs the cloud and fits on the laptop you already own.
24GBthe consumer GPU that now runs a full agent model
Source: Meta AI Research · Tuesday, 11 August 2026