Meta on Monday released Muse Glimmer, its first open-weight AI model in over a year, and CEO Mark Zuckerberg published an essay defending open-source AI, including the right to distill rival models. The 30-billion-parameter model arrives alongside a proposal for a dynamic compute auction, a sketch that raises more questions than it answers.
The release ends a dry spell that began with the spring 2025 flop of Llama 4. That model drew criticism for massaged benchmark numbers, and its largest version never shipped. Now Meta is back with a smaller, agent-focused model and a promise to release its strongest model, Muse Spark 1.2, as open-weight in the coming weeks.
A New Model Built for Local Agents
Muse Glimmer's weights are available under the Apache 2.0 license on Hugging Face. The model is designed for AI agents running locally on a Mac or PC with a single consumer GPU. That is a deliberate shift from the data-center-heavy approach of Meta's earlier models.
At full precision, the model needs over 55 GB of memory. Meta squeezes the weights to about 4 bits, pushing the model under 20 GB. That fits into the memory of current consumer graphics cards and MacBooks. A small helper model speeds up text output by up to 3.1x.
The model was trained by distilling the larger Muse Spark model. That is a common method for building compact models, but it is also the practice that has put Meta at odds with rivals.
Meta compares Glimmer with Google's Gemma4-31B and Alibaba's Qwen3.6-27B. Meta claims Glimmer wins most benchmarks, especially agent tasks like tool use, web search, and long-context work. Qwen is better at driving computer desktop and terminal tasks. On multimodal tasks, all three are even.
Those results come with a caveat. Meta gathered most of the comparison data itself and admits the test setup is not tuned for rival models. Vendor-run benchmarks should be taken with a grain of salt.
The release follows a major reorganization. Zuckerberg rebuilt Meta's AI group as Meta Superintelligence Labs. The company poured billions into Scale AI and poached researchers. Yann LeCun, the former chief scientist, left the company.
Zuckerberg's Essay: Openness as Doctrine
Zuckerberg published an essay titled "The Future is for Everyone." In it, he argues that superintelligence should be spread widely, not held by a few labs. He claims safety comes from balance among many players and that the most dangerous outcome is leading labs keeping the best models to themselves.
The essay also defends distilling other companies' models. That is a direct challenge to OpenAI and Anthropic, which have accused Chinese labs of using their models as teachers without permission. Zuckerberg's principle worth protecting is "that you can learn from anything you can observe."
That line is a clear rebuke to Anthropic CEO Dario Amodei, who has warned about frontier-level open models and pushed for tighter export controls on China. Zuckerberg argues the US should not restrict open models, but should ensure the best open models come from America. He also wants fewer rules on training data for US labs and promises closer cooperation with government, including early model access for safety testing.
The essay reads as an answer to Meta's own investors. Meta trails OpenAI and Anthropic on the most capable models. Open-source strategy is a way to claim leadership where it trails rivals.
The Distillation Fight
Distillation is at the center of the industry's biggest fight. OpenAI and Anthropic see it as theft. Zuckerberg sees it as legitimate learning. The disagreement is not academic.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
In June, Meta limited engineers' use of Claude Code and Codex to avoid training data contamination. An internal memo warned of serious escalations with partner companies. That is a striking move for a company that now argues distillation is fair game.
The tension is real. Meta wants to learn from anything it can observe, but it also wants to protect its own training pipeline from rival tools. The policy suggests Meta sees those tools as a risk, even as Zuckerberg defends the right to use rival models as teachers.
The $600 Billion Question
Meta plans up to $145 billion in investments this year, mostly for data centers. The Wall Street Journal reports Meta plans $600 billion through 2028. That is a staggering number, and there is no direct revenue to match it.
Meta has no API business at comparable scale. Open models bring no licensing money by design. In July, on an earnings call, Zuckerberg floated the idea of a cloud business to monetize data centers. Investors dumped stock after that call.
At an internal town hall, Zuckerberg admitted weaknesses in the AI overhaul. The situation echoes the metaverse. Zuckerberg announced a generational platform shift, renamed the company, and sank tens of billions into Reality Labs without a mass market.
There is a difference this time. AI already improves Meta's core business, including ad targeting and recommendation systems. Demand for AI compute is real, as industry spending shows.
The essay includes a proposal for a "dynamic auction mechanism" for compute. Free versions should reach billions, and anyone wanting more compute pays via auction. Demand sets the price, and scarce data center capacity goes to the highest bidder.
Meta's ad business has run on auction logic for years. That is one of the largest auction machines on the planet. But the compute proposal has no product, timeline, or details on applicability. Whether it becomes a business model worth $600 billion is unanswered.
A Competitive but Not Dominant Position
Meta is competitive but not dominant in small open models. Glimmer is a solid entry, but it is not a clear winner. The company's infrastructure could be the product if it cannot win the model race.
Some observers think Google might follow a similar path after recent upheaval at Deepmind. If Google shifts toward open models, the competitive pressure on Meta increases.
The $600 billion question is whether open models and a compute auction can justify that investment. Zuckerberg's essay argues for openness as a safety measure and a business strategy. His critics see it as a way to avoid the hard work of building a proprietary moat.
The release of Muse Glimmer and the planned open-weight Muse Spark 1.2 are real steps. The essay is a statement of intent. The auction is a sketch. The market will decide if the sketch becomes a blueprint.
For now, Meta has made its bet. It is betting that openness wins, that distillation is legitimate, and that compute can be sold like ads. The next few quarters will show whether investors agree.