Mark Zuckerberg is launching a new AI model and a vision for a more open form of superintelligence.
The social media giant unveiled Muse Glimmer on Monday and said it is an open-weight, agentic model small enough to run on a Mac or PC. Open-weight means users can download and customize the model to run it locally instead of in the cloud.
Meta also plans to release an open-weight version of Muse Spark, its most powerful AI model, soon, Zuckerberg added.
Meta said it trained Glimmer on Muse Spark using a process known as distillation, where a smaller model learns from a larger “teacher” model.
Meta said that Glimmer, the latest model rolled out by the company’s Superintelligence division led by former Scale AI CEO Alexandr Wang, can perform complex reasoning and handle a broad range of agentic tasks while running on a consumer device with a single graphics card.
Glimmer can break a task, such as coding or admin work, into steps. Meta said it “performs strongly for its size class” on several benchmarks for large language models.
‘Filled with doom’
The Meta CEO also published a 6,500-word essay making a positive case for superintelligent AI.
In the essay, titled “The Future Is for Everyone,” Zuckerberg wrote that it was “surprising” that the discourse from so many of the people developing AI was so “filled with doom.”
Zuckerberg said it would be wrong to restrict access to superintelligent AI to a handful of individuals or companies on safety grounds.
“Some argue that superintelligence itself or a small set of experts who control it should decide what is best for humanity. We disagree,” Zuckerberg wrote.
Open vs closed
Meta’s model launch comes as the tech industry is embroiled in a debate over open-weight AI.
Last month, nearly every major AI lab — including Meta — signed a public letter advocating for open-weight models, after the US government suggested it could sanction Chinese open models that were found to have distilled AI models built by their US rivals.
“Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data,” Zuckerberg wrote in his Monday essay. “US policy must reduce this additional friction if we want American open source models to lead over time.”
While Meta has traditionally championed open-sourcing models, not all of the Llama series of open-weight models it first launched in 2023 impressed developers.
Meta took a different approach with Muse Spark, released earlier this year, which was developed under Wang and launched as a closed model. Although the latest version still trails the strongest models from Anthropic and OpenAI overall, it is competitive on several benchmarks.
The decision to launch Glimmer as an open-weight model under a permissive license echoes the strategy of China AI labs, such as Moonshot and DeepSeek, which are launching open models that offer developers a cheaper alternative to their US counterparts.
“Now that Meta Superintelligence Labs are up and running, we will resume releasing some open source models soon,” Zuckerberg wrote in Monday’s blog.
