Meta is back in the open-weight business. Sort of.
On August 10, 2026, Meta Platforms released Muse Glimmer, an AI model that developers can download, modify and run on their own hardware. The company’s most capable system, Muse Spark, stays closed and is available only through an API or paid access.
What Meta actually released
Muse Glimmer has approximately 30 billion parameters. Parameters are the internal settings a model learns during training. More of them usually means more capability, along with heavier hardware demands.
The headline feature is where Glimmer runs. Meta says it works locally on consumer machines, such as a Mac or a PC with a single graphics card. You don’t need a data center or a cloud bill to use it.
Meta built Glimmer for local, agentic work. “Agentic” means the model takes actions for you, not just answers questions. Meta points to tasks like managing calendars, files and messages.
Muse Spark sits on the other side of the wall. It remains Meta’s more advanced model, and it stays proprietary for now. Outsiders can rent its capabilities, but they can’t look under the hood.
That wall may get a door soon. Meta plans to release open weights for a version of Spark, called Muse Spark 1.2, in the weeks after Glimmer’s debut. “Weights” are the trained values that make a model work, so releasing them lets anyone run and adapt that version independently.
The Zuckerberg essay and the market reaction
Mark Zuckerberg paired the launch with a 14-page essay titled “The Future is for Everyone.” Its central argument is that advanced AI capabilities should be widely distributed. He warns against letting a small group of entities hold them.
Investors didn’t seem bothered by that tension. Meta shares rose nearly 3% in premarket trading after the announcement.
How Meta got here
For a while, open weights were Meta’s calling card. Its Llama family made the company the best-known big tech champion of downloadable models. The most recent big entry, Llama 4, came out in April 2025.
Then the direction changed. In 2026, Meta moved toward closed models, and Muse Spark arrived earlier this year as a proprietary system.
Competition explains much of this. Meta is fighting OpenAI, Anthropic and Google at the frontier of AI development. It also faces pressure from Chinese AI labs that have made major advances with open-weight models.
What this means
The most direct effect is on developers. A capable model that runs on one consumer GPU lowers the cost of experimenting. Hobbyists, startups and researchers can build local assistants without paying per query or sending user data to a third party.
The two-track approach also clarifies Meta’s business logic. Glimmer builds goodwill and ecosystem reach, since every developer who adopts it is building on Meta’s foundation. Spark, kept behind an API, gives Meta a product it can charge for and control.
There are risks. Open weights can’t be recalled once they’re published. Anyone can fine-tune Glimmer for purposes Meta never intended.
For investors, the premarket bump is a first reaction, not a verdict. The first signal will be whether Muse Spark 1.2 weights ship on the timeline Meta described. The second is how quickly developers adopt Glimmer compared with open models from Chinese labs.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.