On August 10, Meta released an AI model that anyone can download, modify and use commercially. Muse Glimmer has 30 billion parameters, ships under the permissive Apache 2.0 license and runs on a well equipped home machine once compressed. The more remarkable part is not the model, though, but the 6,500 word essay Mark Zuckerberg published alongside it. In it he defends the very practice that rival labs have called theft.
Key takeaways
- Muse Glimmer is a 30 billion parameter model under an Apache 2.0 license, freely available on Hugging Face.
- At full precision it needs more than 55 gigabytes of memory; compressed to 4 bit it fits into 18 to 20 gigabytes, which brings it onto consumer graphics cards and MacBooks.
- The model was distilled from Muse Spark, Meta’s considerably larger closed model, which stays closed.
- Zuckerberg defends learning from other companies’ models and asks Washington for fewer restrictions on American labs.
- Behind it sits a commitment to spend 600 billion dollars through 2028, with no matching revenue in sight so far.
What the model does
Muse Glimmer is the first open model from Meta Superintelligence Labs, the unit under chief AI officer Alexandr Wang that the company assembled after the Llama 4 family met a cool reception. It is multimodal, trained on more than 100 languages and built for agentic work: tasks where a model does not simply answer, but plans steps, calls tools and checks intermediate results.
Meta compares the model to Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B. By its own account, Glimmer leads on agentic tasks, web research and long contexts, while Qwen does better at driving desktop interfaces and terminal work. A companion model is said to speed up text generation by up to 3.1 times. Vendor numbers deserve caution until independent tests arrive, but the direction is clear enough: Meta is not chasing the top spot in knowledge benchmarks. It wants a model that is useful on hardware you already own.
The real point of contention: distillation
The route to Glimmer ran through distillation, where a smaller model learns from the outputs of a larger one, in this case from Muse Spark, Meta’s closed flagship. The technique is well established but politically charged. When suspicions arose in early 2025 that Chinese labs had trained on the outputs of American systems, the outrage in California was considerable.
Zuckerberg now turns that accusation around. In his essay, titled “The Future is for Everyone,” he defends the principle that you can learn from anything you can observe. The United States, he argues, should not restrict open models but make sure the best open models come out of America. That is at once a counter to OpenAI and Anthropic, both of which have pushed for tighter rules, and a positioning against the Chinese providers who have dominated the open weights market of late.
Where the argument thins out
The contradiction is plain. If you want superintelligence distributed widely, as Zuckerberg writes, you do not release the smaller model and keep the bigger one. Muse Spark remains closed, and what Meta is handing out is the distilled version. An open weight release of Muse Spark 1.2 is planned for the coming weeks, the company says. Until then TechCrunch’s dry observation stands: access is not the same as ownership.
Then there is the arithmetic nobody enjoys doing. Meta has announced 600 billion dollars of investment through 2028, mostly for data centers. Revenue that would justify a number like that does not yet exist. An open model gives money away in the short run. What it buys is developers, attention and political leverage, and it turns your own architecture into the standard others build on. Whether that carries an investment of this size is the actual 600 billion dollar question.
What it means for users
For developers and technically confident users, Glimmer is good news. An Apache 2.0 model can be put to commercial use without asking anyone, and it runs locally. If you cannot or will not hand data to a cloud provider, you now have a serious option on your own hardware. That Meta trained on more than 100 languages matters beyond the footnote it looks like, because smaller language communities benefit from open models more than anyone.
For everyone else, the announcement changes little in the short term. Zuckerberg’s vision of a personal agent working around the clock on your relationships, health, career and finances is a promise, not a product.
Outlook
The interesting question is not whether Glimmer tops the benchmarks. It is whether Meta actually ships the promised release of Muse Spark 1.2. If it does, that would be the largest step yet by an American company towards open frontier models, and Zuckerberg’s essay would gain weight in retrospect. If the smaller model is all we get, the text was mostly a regulatory maneuver on Meta’s own behalf. The coming weeks will settle it.