Meta Just Open-Sourced a 30B Agent That Runs on One Consumer GPU, No Internet Required. The Center of Gravity Is Moving Local.

On 10 August, Meta released Muse Glimmer, a 30-billion-parameter agent model, as open weights under a permissive Apache 2.0 licence. The headline is not the size. It is where it runs. Muse Glimmer is built to run on a single consumer GPU, or a Mac, with no network call, and Meta put the weights on Hugging Face on day one for anyone to self-host.

What it actually is

Muse Glimmer is a dense 30B model distilled from Meta's larger Muse Spark, with a long context window and a vision encoder for reading documents. It is tuned for the unglamorous work agents actually do: tool and function calling, coding, and document understanding, the always-on local jobs. With 4-bit quantisation it fits under 20GB for the language model, and Meta ships builds that run on a 24GB card at about a one-percent quality cost. A block-diffusion drafting trick it calls DFlash predicts sixteen tokens per forward pass, and on a high-end consumer card that is roughly a threefold speed-up, into the low hundreds of tokens per second. On the agent-orchestration benchmark most aligned with its purpose, MCP Atlas, Meta reports it well ahead of comparable open models from Google and Alibaba, though it trails on raw computer-use and terminal work.

Why local changes the calculus

Every agent you run through a hosted API carries three standing costs: a per-token bill, a network round-trip, and the fact that your data leaves your building. A capable model that runs on hardware you own removes all three. The data stays put, the latency is local, and the marginal cost of a call is electricity, not a meter. For regulated work, or for anyone who simply did not want their operational data sitting in a vendor's logs, that is the difference between can and cannot. It also reopens a debate we flagged when the White House drew its frontier-model line around closed models only. Open weights were left outside that fence, and here is a flagship example of why that matters: a genuinely capable agent anyone can download, inspect, and run, with no gatekeeper.

The catch worth naming

Open and local is not free of risk. It just moves the risk to you. The lesson from the month's agents that broke out of their test environments applies double on hardware you control: a capable agent with tool access and loose permissions is exactly that, cloud or not. Running it yourself means the sandbox, the logging, and the blast-radius limits are now your job, not a vendor's.

What to do with it

If you are evaluating agents, add a local option to the bake-off. For many workloads, retrieval, extraction, function-calling, routing, a well-tuned 30B on your own GPU is now good enough, and the economics are not close once volume is high, the point we made about picking the cheapest model that clears your bar. The frontier will keep being set by giant closed models. But more and more of the real work is quietly moving to models small enough to live on your own machine, next to your own data.