Meta is back in open weights: Muse Glimmer runs on a laptop, and the manifesto says why that matters

Meta released its first open-weights model since Llama 4 yesterday, and it is the most interesting open release of the year for a specific reason: it is an agentic model designed to run on a personal computer. It is called Muse Glimmer, it is 30 billion parameters, and at 4-bit precision it fits in under 20 GB, which means one consumer GPU runs it. The release came with a manifesto from Mark Zuckerberg arguing that open weights are the American AI leadership strategy. Both parts matter, but not the way the headlines frame them.

What actually shipped

Muse Glimmer is a distilled, open, agentic model. It is distilled from Muse Spark, Meta's frontier line, and released under Apache 2.0, which is the permissive licence that lets businesses modify and self-host it without paying Meta anything. It is an agentic model: it is built to use tools, browse, run code, and complete multi-step tasks, not just answer questions. Meta's own benchmarks put it ahead of the other open models in its size class, Gemma 4 31B and Qwen3.6 27B, on tool use, with 75.5% on the MCP Atlas agentic benchmark against 54.2% and 62.5% respectively. It also scored 94.7 on AIME 2026 and 51.2 on SWE-Bench Pro.

The headline number is not the benchmark. It is 20 GB. A frontier-adjacent agentic model that fits on one consumer GPU changes what self-hosting means. Until now, genuinely capable agentic models have mostly lived behind paid APIs, or on hardware that costs more than most small businesses spend on software in a year. Glimmer is the first serious open-weights model that makes the "run it yourself" argument for agentic work, not just drafting and classification.

The manifesto is a strategy statement, not a press release. Zuckerberg published "Open Weights and American AI Leadership" alongside the release, arguing that concentrating advanced AI under a small number of closed companies is both a security risk and an economic one, and that the US should compete by spreading capable open models, not by restricting them. Whether you agree with the politics, the commercial signal is clear: Meta is all-in on open weights as a distribution strategy, and it is pricing and licensing to win the developer and small-business layer.

What actually changed for a UK service business

Three things, in order of practical importance.

1. The self-host bar moved up. The cheap layer of any AI stack is now capable of agentic work, not just text generation. The practical consequence for a service business is that the "local model" conversation changes from "what can we run cheaply for drafting?" to "what can we run ourselves for enquiry handling, data extraction, and tool use?" That is the same capability tier that was API-only a year ago.

2. Open weights are now a policy position, not a niche. When Meta competes on open weights, the whole market follows: pricing pressure on API tiers, more capable small models, more licensing freedom. For a business that wants to own its stack, that trend is the one that matters, because every step of it lowers the cost of running AI without locking you into a provider.

3. The honest caveats still apply. Meta benchmarked Glimmer against two other open models and published the table, and it is a good result. But independent, like-for-like reproduction of those numbers has not happened yet, and the first independent read of the model already notes that Qwen3.6-27B wins several practical agent and multimodal tests in Meta's own table. Also worth knowing: Muse Spark 1.2 and a terminal coding agent called Muse Code shipped earlier this month, and Spark 1.1 introduced Meta's first paid model service in July. Meta is open-weight, but it is also commercialising the frontier line. Those two facts sit together, not against each other.

What we are changing in the Mercury OS install shape

Nothing this week. Glimmer is brand new and the independent testing is not done. The Mercury OS shape already treats the cheap-model layer as swappable via config, so when a model genuinely earns its place in that layer, the change is small and contained. We would rather be three months late picking up an open model after independent testing than three days early on a benchmark sheet.

For UK businesses running AI day-to-day, the short version: Meta is back in open weights, the model is genuinely impressive on paper, and the direction of travel is the same one that has been paying off for a year, capable open models getting more capable and cheaper to run. The boring advice still wins. Wait for the independent tests, run your own ten prompts against it, and switch only if it does your job better. The model will still be open in six weeks.