After more than a year off, Meta is releasing model weights again. The compact agent model Muse Glimmer holds its own against rivals, and Mark Zuckerberg paired it with an essay that reads like a reply to OpenAI and Anthropic.

Meta has released Muse Glimmer, a 30-billion-parameter model with weights available under an Apache 2.0 license on Hugging Face. It's built for AI agents that run locally around the clock on a Mac or PC with a single consumer GPU. It's Meta's first open model since Llama 4 in spring 2025. And it won't be the last. According to the Wall Street Journal, Meta plans to ship an open-weight version of Muse Spark 1.2, its strongest model right now, in the coming weeks.

The stretch in between was turbulent. Meta was long the most important source of open models thanks to Llama, but Llama 4 flopped and drew criticism over massaged benchmark numbers. The largest version of the family never shipped at all. Zuckerberg rebuilt the AI group as Meta Superintelligence Labs, poured billions into data provider Scale AI and into poaching top researchers, and reorganized the unit several times. Chief scientist Yann LeCun, the face of Meta's AI research for years, left the company. Muse Glimmer is the new unit's first open model.

Meta is competitive, but not dominant

Meta compares Glimmer with Google's Gemma4-31B and Alibaba's Qwen3.6-27B, the two leading open models in this size class. Glimmer wins most benchmarks, especially on agent tasks like tool use, web search, and long-context work. Qwen is clearly better at driving a computer desktop and at terminal tasks. On multimodal tasks, all three are even.

So Meta is back in the game among small open models, but not ahead of it. And as always with vendor-run benchmarks, take the numbers with a grain of salt. Meta gathered most of the comparison data itself and admits in its methodology report that its test setup isn't tuned for the rival models.

Built to run on your own hardware

At full precision, the model would need more than 55 GB of memory. Meta squeezes the weights down to about 4 bits, pushing the model under 20 GB. That's small enough to fit, image processing included, into the memory of current consumer graphics cards and MacBooks. An small helper model speeds up text output by up to 3.1x, Meta says. The pitch: an agent that handles your calendar, files, and personal messages should run entirely on your device, with no data going to a cloud.

Glimmer was trained by distilling Meta's larger Muse Spark model, meaning the small model learns to copy the big one's outputs. That's how most compact models get built today.

Zuckerberg defends distilling other labs' models, Amodei calls it a threat

Alongside the release, Zuckerberg published an essay titled "The Future is for Everyone" laying out the case for Meta's open-source strategy. His main argument is that superintelligence shouldn't sit with a handful of labs but should be spread as widely as possible. A single benevolent superintelligence can't exist, he writes, and safety comes from a balance among many players. The most dangerous outcome would be leading labs keeping their best models to themselves.

What stands out is how bluntly he defends distilling other companies' models. Some try to paint it as harmful, he writes, but the principle worth protecting is "that you can learn from anything you can observe." That lands in the middle of an ongoing fight. OpenAI and Anthropic have repeatedly accused Chinese labs of using their models as teachers without permission, and Anthropic CEO Dario Amodei has warned for years about frontier-level open models while pushing for tighter export controls on China. Distillation uses frontier model outputs to generate training data for various training stages. Companies like OpenAI see it as a free ride for Chinese labs and effectively theft, given what they spend on their own training pipelines.

Zuckerberg takes the opposite view. The US shouldn't restrict open models, he argues, but make sure the best open models come from America, distillation included. He also wants fewer rules on training data for US labs, and in exchange promises closer cooperation with the government, such as early model access for safety testing.

There's a business reason Meta of all companies makes this argument. On the most capable models, it trails OpenAI and Anthropic. Open, widely distributed models are the one area where Meta could plausibly claim the lead. Right now Chinese labs fill that role, the same labs OpenAI and Anthropic accuse of feeding on their model outputs.

Just how touchy the subject is became clear only weeks ago. In June, according to The Information, Meta limited how its own engineers could use Anthropic's Claude Code and OpenAI's Codex, so their outputs wouldn't end up in Meta's training data. An internal memo warned of serious escalations with partner companies. Rhetorically, at least, that escalation has now happened.

The open question: what do investors get for $600 billion?

For shareholders, a bigger question looms. Meta plans up to $145 billion in investments this year alone, mostly for data centers, and $600 billion through 2028, per the WSJ. So far there's no direct revenue to match. Meta trails OpenAI and Anthropic on top-end models, there's no API business at a comparable scale, and open models like Glimmer bring in no licensing money by design.

July showed how impatient Wall Street has gotten. On the earnings call, Zuckerberg floated the idea of a cloud business that would monetize Meta's data centers directly. He gave no details, and investors dumped the stock. At an internal town hall, he also admitted weaknesses in the company's AI overhaul.

The metaverse parallel is hard to miss. Back then he announced a generational platform shift, renamed the company, and sank tens of billions into Reality Labs over the years without a mass market ever showing up. One difference matters, though. AI already improves Meta's core business, including ad targeting and recommendation systems, and demand for AI compute is real, as the whole industry's spending shows. But what if Meta builds enormous infrastructure for models it then gives away while rivals sell their best ones?

A hint at how Zuck plans to make money hides in a subordinate clause of the essay. Free versions should reach billions of people, and anyone who wants more compute pays for it through a "dynamic auction mechanism." Demand sets the price, and scarce data center capacity goes to whoever will pay the most.

Meta knows this playbook well. Its ad business, one of the largest auction machines on the planet, has run on the same logic for years. Zuckerberg would port that pricing model to compute, which also fits the cloud ideas he hinted at in July. But it's still just a sketch. No product, no timeline, no word on whether it applies to consumers, developers, or enterprises.

So the essay also reads as an answer to Meta's own investors. If Meta can't win the race for the best model, it declares distribution the real goal and its infrastructure the product. Some observers think Google might follow a similar path after the recent upheaval at Deepmind. Whether that turns into a business model worth $600 billion is a question Zuckerberg's philosophy can't answer on its own.

Read on for the full picture.
Subscribe for hype-free coverage.

• Access to all THE DECODER articles.
• Read without distractions – no Google ads.
• Access to comments and community discussions.
• Weekly AI newsletter.
• 6 times a year: “AI Radar” – deep dives on key AI topics.
• Up to 25 % off on KI Pro online events.
• Access to our full ten-year archive.
• Get the latest AI news from The Decoder.