Meta Superintelligence Labs: what does Meta mean by superintelligence?

This page is part of our series on who is building superintelligence. Its main job is to explain what Meta means when it uses the word.

What Meta Superintelligence Labs is

On 30 June 2025 Mark Zuckerberg told staff in an internal memo, later published in full by the press, that Meta would call its “overall organization Meta Superintelligence Labs (MSL)”. It brought together Meta’s foundation-model teams, its product teams and its long-standing research group FAIR, together with a new laboratory. Alexandr Wang, previously chief executive of Scale AI, was appointed chief AI officer to lead it, and Nat Friedman, a former GitHub chief executive, to head AI products and applied research. The memo gave the aim as “personal superintelligence for everyone”.

The unit has been reorganised several times since, according to press reports: split into four groups in August 2025, with about 600 jobs cut in October 2025. Yann LeCun, Meta’s chief AI scientist and FAIR’s founder, announced in November 2025 that he would leave at the end of the year to start his own company. Daniel Gross, a co-founder of Safe Superintelligence Inc., joined MSL in mid-2025 (see our SSI profile). Meta has not published a primary account of most of these changes.

What Meta means by “superintelligence”

The fullest statement of Meta’s view is a letter Zuckerberg published on 30 July 2025 titled “Personal Superintelligence”. Its central idea is “everyone having a personal superintelligence that helps you achieve your goals, create what you want to see in the world” and be a better friend to those you care about. He contrasted this with a vision of AI “directed centrally towards automating all valuable work”, after which, he wrote, “humanity will live on a dole of its output”.

The letter also makes stronger claims. It says that “developing superintelligence is now in sight”, that Meta has seen “glimpses of our AI systems improving themselves”, and that “the rest of this decade seems likely to be the decisive period” for the technology. It predicts that glasses that understand their wearer’s context “will become our primary computing devices”.

What the letter does not do is define superintelligence. It describes it through what it would be used for. Meta’s later statements follow the same pattern: its April 2026 safety framework says Meta’s vision is “to bring personal superintelligence to everyone”, and its model announcements describe each release as a step towards it.

How this differs from the research sense

	Superintelligence in research (Bostrom, 2014)	“Personal superintelligence” (Meta, 2025)
What it is	A system that greatly exceeds the cognitive performance of humans in virtually all domains	A personal AI assistant or agent that helps an individual with their goals
How it is defined	By capability relative to the best humans	By its use and its relationship to the user; no capability threshold given
Who it serves	Not specified; the definition is about the system	Each individual user, through Meta’s apps and devices
Status	Hypothetical; nothing meets the definition	Described by Meta as a goal; current models presented as early steps

The two can overlap: a sufficiently capable personal assistant might also be superintelligent in the research sense. But they are answers to different questions. The research sense asks how capable a system is; Meta’s phrase describes what kind of product it wants to offer. A reader who sees “Meta is building superintelligence” should not assume it means the same thing as a researcher warning about superintelligence. Our explainer covers the research meaning, and AGI vs superintelligence explains why the word now has several meanings in circulation, including the US administration’s use of “Super Intelligence” for AI in general.

What it has built

Before MSL, Meta’s flagship models were the Llama family, released with downloadable “open weights”. Llama 4, released in April 2025, was received less well than Meta had hoped, and Meta was criticised for how one version was presented on a public leaderboard.

Under MSL, Meta has released a new family called Muse:

• Muse Spark (8 April 2026), MSL’s first model, described by Meta as a “first step toward a personal superintelligence”. Meta says it can reach Llama 4 Maverick’s capabilities with over an order of magnitude less compute. Meta’s announcement does not offer its weights for download.
• Muse Image and Muse Video (announced 7 July 2026); Meta said Muse Video was “coming soon”.
• Muse Spark 1.1 and a Meta Model API in public preview (9 July 2026).
• Muse Spark 1.2 and Muse Code, a coding agent in beta (5 August 2026).
• Muse Glimmer (10 August 2026), a smaller model released with open weights under the Apache 2.0 licence for agents running locally.
• Muse Spark 1.3 (2 September 2026), which Meta says “advances our work toward personal superintelligence”. Meta lists an open-weights release of Muse Spark on its roadmap, without a date.

On its July 2026 earnings call, Zuckerberg said Meta had “always basically said that we were going to do a mix of open and closed”, though Meta had previously championed open weights for its Llama models. Press reports in late August 2026, citing internal documents, said Meta was preparing a consumer agent platform called Hatch and a further model; Meta declined to comment, and as of 3 October 2026 this site had not confirmed a launch.

Meta has not presented any MSL model as superintelligent, and this site found no independent evaluation placing an MSL model ahead of the leading systems from other developers. The benchmarking firm Artificial Analysis placed Muse Spark 1.3 close behind them in September 2026.

Infrastructure

Meta is spending heavily on the computing power its ambitions require. In its July 2026 results it put 2026 capital expenditure at $130 billion to $145 billion, including finance-lease payments. Zuckerberg said in July 2025 that its Prometheus data centre in Ohio would come online in 2026 at about one gigawatt, and that a larger site, Hyperion in Louisiana, would eventually reach five gigawatts. Those are plans and statements; how much of the spending goes to MSL’s work specifically is not published.

Safety and governance

Zuckerberg’s letter said that superintelligence “will raise novel safety concerns” and that Meta would need to be “careful about what we choose to open source”. Meta’s formal policy was its Frontier AI Framework of February 2025, revised and renamed in April 2026 as the Advanced AI Scaling Framework. The 2026 version says Meta will proceed with developing a model that reaches its highest risk tier “only if catastrophic risks are within acceptable levels”, and will deploy only with mitigations that have been “defined, implemented and validated”. Its evaluations are Meta’s own, supplemented by outside experts it engages.

Independent assessments are critical. The Future of Life Institute’s AI Safety Index for summer 2026 gave Meta a D+ overall and an F for existential safety, though it noted improvement on the previous edition. That index is one organisation’s rating, not a regulatory finding. On 29 September 2026 Zuckerberg signed the voluntary White House Accord on Super Intelligence, which commits signatories to internal monitoring and an internal oversight team, an independent external auditor or evaluator, and an independent board committee to oversee their most capable models.

The general questions these policies try to answer, how to keep advanced systems aligned and correctable, are covered on our pages on alignment and control.

Ambition and evidence

• Stated ambition: personal superintelligence for everyone; superintelligence “in sight”; AI glasses as the main computing device; delivering personal superintelligence to billions of people through Meta’s products.
• Demonstrated: a new family of models released during 2026, a mix of closed and open weights, and very large infrastructure spending.
• Not established: any capability threshold for what Meta would call superintelligence, or independent evidence that its systems are approaching superintelligence in the research sense.