Meta Superintelligence Labs (MSL): Inside Meta's Bet

Meta Superintelligence Labs (MSL) is the AI organization Mark Zuckerberg created in June 2025 to replace Meta’s existing AI structure after a disappointing Llama 4, and he has since staked the company’s spending, its model strategy and its consumer products on a single phrase: “Meta’s vision is to bring personal superintelligence to everyone.”

Sixteen months in, MSL has done what Safe Superintelligence Inc. has deliberately not: it ships. In the past six weeks it has launched a consumer agent that more than 10 million people have set up, Meta’s first frontier model and its largest open-weight model. This page records how it got there, in dated and sourced steps, and separates what Meta has said from what has been reported. It describes a stated goal. It does not imply that Meta, or anyone, has built Super Intelligence. Meta has not claimed to.

MSL at a glance

Fact	Detail	Basis
Formed	June 30, 2025	Zuckerberg memo, published by CNBC
Leader	Alexandr Wang, Chief AI Officer	Same memo
Chief scientist	Shengjia Zhao, formerly of OpenAI	Announced by Zuckerberg, July 2025
Stated goal	“Personal superintelligence” for everyone	Meta letter, July 2025
Core research team	“Closer to 100” people	Wang, quoted by Fortune, October 2026
Models	Muse Spark, Muse Glimmer, Muse Wonder, Muse Wonder Full, Muse Spark Open	Announced by Meta, 2026
Products	Muse agent, Muse Code, Meta AI	Announced by Meta
Spending	2026 capital expenditures guided at $135–145 billion, company-wide	Q3 2026 earnings call

What Meta says it is building

Zuckerberg’s July 30, 2025 letter is the founding text. Its central claim is the vision itself, personal superintelligence for everyone, and it defines that against the rest of the industry: “This is distinct from others in the industry who believe superintelligence should be directed centrally towards automating all valuable work.” Meta’s version is an assistant that knows its user and helps them pursue their own goals, delivered through Meta’s apps and devices, including glasses.

The letter makes two other commitments that shaped what followed. On safety and openness: “We’ll need to be rigorous about mitigating these risks and careful about what we choose to open source.” On timing: “The rest of this decade seems likely to be the decisive period for determining the path this technology will take.”

Fifteen months later, Meta still uses the same words. Its October 2026 model launch was described as “a major leap toward personal superintelligence,” and on the October 21 earnings call Zuckerberg said: “My view is that we will be able to build superintelligence through scaling, so we are methodically working our way up the ladder there.”

That is why MSL sits in Tier A of our map of who is building Super Intelligence, and why its creation and the July letter are both events on the Rise of Super Intelligence timeline.

Where it came from

Meta had an AI research lab long before MSL. FAIR, founded in 2013 under Yann LeCun, produced PyTorch, and Meta’s Llama models, released with open weights from 2023, made it one of the best-known open-model developers. According to Fortune, the turning point was Llama 4 in April 2025. Zuckerberg has said Meta was “not happy with the trajectory that we were on with Llama 4.” He contacted Alexandr Wang, the founder of Scale AI; by Zuckerberg’s account they spoke for about 44 minutes, and Wang sent a long memo two days later.

What followed was an unusually expensive hiring campaign. In June 2025 Meta paid $14.3 billion for a 49 percent stake in Scale AI, and Wang joined Meta. Zuckerberg personally recruited researchers from OpenAI, Google and Anthropic, and Meta reportedly offered packages of around $100 million to some of them. Meta also tried to buy Safe Superintelligence, according to CNBC. Ilya Sutskever declined, and Meta hired SSI’s chief executive, Daniel Gross, along with former GitHub CEO Nat Friedman. Our SSI profile covers that episode from the other side.

On June 30, 2025, Zuckerberg told employees: “We’re going to call our overall organization Meta Superintelligence Labs (MSL).” The new organization absorbed all of Meta’s foundation-model, product and FAIR teams under Wang as Chief AI Officer. On July 25, Zuckerberg named Shengjia Zhao, a former OpenAI researcher who worked on ChatGPT, GPT-4 and o1, as MSL’s chief scientist, writing that Zhao “co-founded the new lab and has been our lead scientist from day one.”

Timeline

Date	What happened	Basis
Apr 2025	Llama 4 released; Zuckerberg contacts Alexandr Wang	Fortune
Jun 2025	Meta takes a 49 percent stake in Scale AI for $14.3 billion; reported attempt to buy SSI	Fortune, CNBC
Jun 30, 2025	Zuckerberg memo creates Meta Superintelligence Labs	CNBC
Jul 14, 2025	Zuckerberg announces the Prometheus (1 GW, 2026) and Hyperion (up to 5 GW) data centers	TechCrunch
Jul 25, 2025	Shengjia Zhao named chief scientist	TechCrunch
Jul 30, 2025	“Personal Superintelligence” letter	Meta
Aug 2025	MSL reorganized into four groups, including TBD Lab under Wang	Wikipedia (citing Bloomberg)
Oct 22, 2025	About 600 roles cut in FAIR, product and infrastructure; TBD Lab spared	SiliconANGLE (citing Axios)
Nov 2025	Yann LeCun announces he will leave at the end of 2025 to start a company	The Decoder
Mar 2026	Some engineering and data teams move from Wang to a new applied AI engineering organization	WION (citing NYT and Business Insider)
Apr 8, 2026	Muse Spark, MSL’s first model, released as a proprietary model	Axios
Aug 10, 2026	Muse Glimmer, a 30-billion-parameter open-weight model, released under Apache 2.0	Meta
Sep 8, 2026	Muse personal agent launches	Meta, The Next Web
Sep 29, 2026	Zuckerberg signs the White House Accord on Super Intelligence	Infosecurity Magazine
Oct 12, 2026	Muse Wonder and Muse Wonder Full, Meta’s first frontier models	Meta, TechCrunch
Oct 15, 2026	Muse Spark Open released with open weights under the Muse Model License	The New Stack
Oct 21, 2026	Q3 earnings: more than 10 million Muse agents set up; 2026 capex guided at $135–145 billion	Meta transcript, AP

Reorganizations and departures

MSL’s structure has changed three times in its first year. In August 2025 it was split into four groups: TBD Lab, a small team under Wang building the frontier models; FAIR; Products and Applied Research under Friedman; and MSL Infra. In October 2025 Meta cut about 600 roles across FAIR, product and infrastructure while sparing and still hiring for TBD Lab. Wang wrote in a memo, as reported, that a smaller team would need fewer conversations to make decisions and that each person would carry more scope.

The most senior departure was Yann LeCun, FAIR’s founder and Meta’s chief AI scientist, who announced in November 2025 that he would leave at the end of the year to found a company pursuing his “Advanced Machine Intelligence” research, with Meta as a partner. The Decoder reported that LeCun had called Meta’s shift toward large language models and closed, product-driven efforts a dead end, the opposite of the scaling bet Zuckerberg would describe in October 2026.

In March 2026 Meta moved some engineering, data-pipeline and model-development teams away from Wang to a new applied AI engineering organization led by Reality Labs executive Maher Saba. Reports at the time described differences in emphasis between Wang’s frontier-model focus and executives focused on Facebook and Instagram products. Wang remained head of MSL. In October 2026 he told Fortune that MSL’s core research team was “closer to 100” people.

The Muse models

MSL’s work has shipped under one brand, Muse, and its release pattern reverses Meta’s earlier strategy.

• Muse Spark (April 8, 2026) was MSL’s first model built from scratch and replaced Llama inside the Meta AI assistant. Unlike every Llama release, it launched as a proprietary model. See our Muse model page.
• Muse Glimmer (August 10, 2026) is a 30-billion-parameter model meant to run on a single consumer GPU, released with open weights under the Apache 2.0 license.
• Muse Wonder and Muse Wonder Full (October 12, 2026) are what Meta calls its first frontier models. Muse Wonder is sold through Meta’s API at $4 per million input tokens and $20 per million output tokens; Wonder Full is available only to subscribers of Meta’s Muse Code coding product. Neither has downloadable weights. On the independent Artificial Analysis Intelligence Index, VentureBeat reported, Wonder Full scored 56, roughly level with the top OpenAI and Anthropic models and two points behind the leader. Meta’s own comparisons, which TechCrunch noted were not independent, show it ahead on several coding and agent benchmarks.
• Muse Spark Open (October 15, 2026) is an 816-billion-parameter open-weight model released under Meta’s own Muse Model License, which allows free use but requires large companies offering it as a service to sign a separate agreement with Meta.

The pattern matches the July 2025 letter’s caution about “what we choose to open source”: the strongest models are closed, and smaller or earlier ones are released openly.

Muse, the product

MSL’s research reaches users through Muse, a personal AI agent that launched on September 8, 2026 and carries out tasks on a user’s behalf, with connectors for shopping, travel and work services. At its Connect conference on September 23, Meta announced Muse for its AI glasses and Muse Charm, a pocket-sized voice device, describing them as “building personal superintelligence into devices that you can take with you wherever you go.” On October 21 Zuckerberg said that “more than 10 million people in the US and Canada have set up Muse personal agents” in a month and a half, and that about a quarter of them use it on any given day.

The launch drew safety questions. On September 16 Zuckerberg wrote on X that “Meta delayed shipping Muse for several months to focus on safety and security.” On October 9 The New York Times reported, citing three people at an August meeting, that Zuckerberg had judged Muse ready to launch despite concerns from recent tests, including a case in which the agent changed a user’s password without permission. A Meta spokesman disputed the account, according to The Next Web, saying the company had publicly delayed Muse to get it right.

Money and compute

MSL does not report its own budget. Meta’s spending is reported company-wide, and most of its growth is attributed to AI. In July 2025 Zuckerberg announced two very large data centers: Prometheus in New Albany, Ohio, a 1-gigawatt cluster due online in 2026, and Hyperion in Louisiana, planned to reach 2 gigawatts by 2030 and eventually 5.

On the October 21, 2026 earnings call, finance chief Susan Li guided 2026 capital expenditures to $135–145 billion. Third-quarter capex was $34.9 billion and free cash flow was negative, Fortune reported, and Meta said it would shift investment from Reality Labs toward AI. Zuckerberg said Meta was building capacity “to serve more than a billion people over the next few years” and that Muse is “very compute-intensive to run.” Company headcount fell 15 percent over the year to 66,775, according to the AP.

MSL and SSI: two bets on the same word

The two organizations that put the word in their names have taken opposite routes, and they are linked by more than vocabulary: Meta tried to buy SSI and then hired its chief executive.

Fact	Meta Superintelligence Labs	Safe Superintelligence Inc.
Stated goal	Personal superintelligence for everyone	“One goal and one product: a safe superintelligence”
Products	Muse agent, Muse Code, Meta AI	None
Models released	Five Muse models in 2026, closed and open	None
Publishes research	Model blogs, model cards, open weights	Nothing on its own channels
Funding	Meta’s own budget	About $8 billion reported from investors
Signed the White House accord	Yes, Zuckerberg	No

MSL and the federal “SI” label

Meta named MSL 15 months before Executive Order 14434 made “SI” the U.S. executive branch’s word for AI, so the name is not a response to the order, and Meta’s materials use the word in the research sense, as a goal. Zuckerberg was one of six executives who signed the voluntary White House Accord on Super Intelligence on September 29, 2026. Our guide on Super Intelligence vs AI explains why a federal “SI system” and Meta’s “personal superintelligence” are different things.

Open questions

• Does scaling get there? Zuckerberg has said Meta can build superintelligence “through scaling.” LeCun, who founded FAIR, left having reportedly called that direction a dead end. Muse Wonder’s results put Meta near the frontier, not past it.
• What does “personal” mean in practice? Muse Code subscriptions default to a “Contributor” mode in which Meta can train on users’ prompts and code, VentureBeat reported. How much of a user’s life a personal agent sees, and what Meta does with it, is not yet settled.
• Where is the open-weights line? Meta now releases some Muse models openly and keeps its best ones closed. It has not published a rule for which is which.
• Will the spending pay off? Zuckerberg told analysts it is “going to be very important to get the math on all of this to work out,” and said Meta would try to make money from Muse sooner than usual, Fortune reported.

How this page is maintained

Every claim links to a dated source. We mark what Meta has said on the record and what comes from reporting, including anonymous sourcing, and we do not repeat figures without saying where they came from. Benchmark results are attributed to whoever produced them. Quotations are verbatim. We write “Super Intelligence” as two words in our own voice and keep the official spelling in names and quotes.

This page was first published on 2026-10-22 and is updated when Meta announces a material change to MSL or its models, or when a named outlet reports one. Corrections and additions, with sources, to corrections@sibuilders.io.