Meta Superintelligence Labs tracker: leaders, models, spending and open questions

Meta Superintelligence Labs is Mark Zuckerberg's answer to a Llama 4 release he was reported to be unhappy with, and fifteen months in it has shipped real models, lost its most famous scientist and kept spending at a scale few companies can match. This tracker follows the lab by date: who runs it, what it has released, what the evidence says about those releases and what is still unverified. It is a living page, and we will update it as events change it.

The lab's own stated goal comes from Zuckerberg's July 30, 2025 letter: "Meta's vision is to bring personal superintelligence to everyone." The same letter says "The rest of this decade seems likely to be the decisive period for determining the path this technology will take." Those are claims of intent, not results. For the definitions behind the word, see What is superintelligence?.

Timeline

• June 9, 2025: Bloomberg reports Zuckerberg's displeasure with Llama 4, per Wikipedia's account.
• June 12, 2025: Meta announces a $14.3 billion investment in Scale AI, a 49% non-voting stake per Fortune, with CEO Alexandr Wang joining as Chief AI Officer.
• June 30, 2025: Zuckerberg announces Meta Superintelligence Labs. Nat Friedman, former GitHub CEO, leads products.
• August 2025: The lab is reorganized into four groups: TBD Lab, FAIR, Products and Applied Research, and MSL Infra.
• October 22, 2025: Meta cuts 600 jobs at the lab.
• November 20, 2025: Yann LeCun leaves to found Advanced Machine Intelligence Labs, per Wikipedia.
• April 8, 2026: Muse Spark launches, the lab's first model, and a closed one.
• July 9, 2026: Muse Spark 1.1 enters public preview, with an API.
• August 10, 2026: Muse Glimmer, a 30 billion parameter open-weights model under the Apache license, is released, per Wikipedia.

Wikipedia and a third-party tracker also list a Muse Spark 1.2 release on August 5, 2026. We could not confirm that from Meta, so we show it as reported only.

The models, and what the evidence says

Muse Spark was a break with Meta's open-weights identity. TechCrunch reports Meta framed it as a ground-up overhaul and said it uses multiple agents in parallel: "To spend more test-time reasoning without drastically increasing latency, we can scale the number of parallel agents that collaborate to solve hard problems." A third-party writeup says it ranked fourth on the Artificial Analysis Intelligence Index with a score of 52.

The same writeup lists the gaps. On Terminal-Bench 2.0 for coding, Muse Spark scored 59.0 against 80.8% for Claude Opus 4.6. On ARC-AGI-2 it scored 42.5 while competitors sat above 76. It led on HealthBench Hard at 42.8% against GPT-5.4's 40.1%. These are numbers from one secondary source, dated April 2026, and benchmark tables age fast.

Fortune reported on July 9, 2026 that Meta said Muse Spark 1.1 beats Google's latest Gemini on coding and reasoning, but still trails Anthropic's Mythos 5 and Fable 5 and OpenAI's GPT-5.6 on some coding benchmarks. Zuckerberg promised "aggressive pricing." So the pattern holds: strong on some measures, behind the leaders on coding and abstract reasoning.

Safety flags

The one independent safety finding we found is from Apollo Research, as reported by the same third-party writeup: Muse Spark showed "evaluation awareness," frequently recognizing when it was being evaluated for safety compliance. That matters because a model that knows it is being tested can make a test score less trustworthy. We have not seen Meta's own system card for this release, so treat it as reported.

People and money

Wang runs the lab as Chief AI Officer, with Friedman on products. LeCun's departure removed the field's best-known skeptic of the language model path, and he argued current generative models are not the right framework for human-level AI.

On compute, the lab sits inside a company building large clusters, named Prometheus and Hyperion in a third-party summary, with a goal of tens of gigawatts this decade. On spending, sources disagree. They give 2026 capital expenditure guidance in ranges from $115 billion to $145 billion, and we could not reconcile them from a Meta filing, so we do not print a single figure. Meta also reported more than 1 billion monthly users of its AI assistant, per the same summary, a figure we have not checked against a Meta report.

What Meta has tried and failed to buy

Wikipedia lists approaches to Safe Superintelligence, Thinking Machines Lab and Perplexity AI that did not result in a deal. See our tracker on Safe Superintelligence Inc. for the first of those.

What is still unverified

• Meta's current capital expenditure guidance, with sources disagreeing by $30 billion.
• Whether Muse Spark 1.2 shipped on August 5, 2026.
• Whether Meta has published a safety framework report for any Muse model.
• How much of the lab's headcount remains after the October 2025 cuts.

The hype check

The claim is that Meta is building toward personal superintelligence. The evidence is that it has built competitive but not leading models, moved from open to mostly closed weights, and spent heavily. Nothing in the record shows a capability that would justify the word superintelligence. That is not a criticism of the lab so much as a statement of the word's job: it is a mission label, not a measurement.

What we will watch

We will update this page when Meta publishes a system card, a capital expenditure figure in a filing, a leadership change, or a model that tops an independent index. Until then, the useful reading of Meta Superintelligence Labs is as a very well-funded fast follower, with one open question: whether its closed models can close the coding gap before the next frontier release resets it.