Meta open-sourced a 30-billion-parameter agent model that runs on a Mac or a single consumer GPU, which moves continuous AI work from a metered bill to a fixed cost you already paid.
Think about the agent you have wanted to run and haven't: the one that sits on your support queue all day, reads every ticket as it lands, tags it, drafts a reply, and escalates the three that actually need a person. Nobody skips that project because the model can't do the work. They skip it because "all day, every day" is a billing model, and the meter runs whether or not anything interesting happens. On Monday, Meta open-sourced a model that does that job on a computer you already own, at a per-token cost of zero.
Muse Glimmer is a 30-billion-parameter model that Meta Superintelligence Labs released on August 10 under an Apache 2.0 license, the permissive kind you can use commercially without asking anyone. Per Meta's research blog, it was distilled from the larger Muse system and built for agent work rather than conversation: reliable tool calling, multi-step reasoning, recovering when a step fails, and holding state across sessions that run for hours. It reads images through a small perception encoder attached to the text decoder, and it handles more than 100 languages.
The commercially interesting part is the hardware floor. Meta validated it on a MacBook M4-Max, an M5-Max, and an RTX-5090. Quantized, it fits in 24 to 32 GB of memory. Simon Willison ran an 18 GB build through LM Studio, pointed it at a real codebase, asked how authentication worked, and it went and found out. Weights are on Hugging Face, with support arriving across Ollama, vLLM, and the usual hosted providers.
Here is the reframe. For two years the practical question about any AI workflow has been "can the model do this," and the answer kept improving while the bill got worse in one specific way: the more continuous the work, the worse the economics. On-demand tasks are cheap because they are rare. Always-on tasks are expensive because they are always on. So teams put AI behind buttons a person clicks, and left the watching, monitoring, and waiting jobs alone.
A model small enough to live on a machine in your office changes which side of that line a project lands on. Artificial Analysis lists Muse Glimmer's price as $0.00 per million tokens, for the obvious reason that you are the one running it. Once inference is a fixed cost you have already paid, "how often should this run" stops being a budget question and becomes an engineering one. Run it every minute. Run it against every record. Run it on the boring 95 percent of tickets that were never worth an API call, which is exactly the tier where the unglamorous savings live.
Open weights have mostly been a portability story up to now. You got an exit from your vendor, not a smaller bill, because you still needed a data center to run the thing. This one is different in the only way that matters to a budget owner: the floor is a laptop. And the data never leaves the building, which for anyone who has sat through a compliance conversation about pasting customer records into a vendor API is not a cost story at all. It is a permission story.
Now the part the launch posts skip. Muse Glimmer is not going to beat the model you are currently paying for. The Register notes it is too small to compete with the Chinese open-weight heavyweights like Qwen 3.8-Max and Kimi K3, and clocks real-world throughput anywhere from 6 to 233 tokens per second depending on what you run it on. That bottom number is not a typo, and on the wrong hardware it turns an always-on agent into an always-behind one. Artificial Analysis ranks it second in its class, and the phrase carrying the weight there is "in its class," meaning open-weight models of similar size, not the frontier. Willison's own testing found it capable on a codebase question and jumbled on a creative one.
Read the positioning honestly and this is a small-to-medium business tool: good enough for the repetitive tier of work, not the judgment tier. Which is fine, because the repetitive tier is where almost all of the unclaimed automation actually sits. The projects nobody built were never blocked on intelligence. They were blocked on the fact that watching something costs the same as thinking about it.
Frontier models made AI capable of more. This one makes it cheap enough to be constant, and constant is the setting most businesses have never once been able to try.