Meta Superintelligence

Meta Superintelligence Labs (MSL) is Meta’s AI organization behind the Muse model family. Its work spans multimodal reasoning, tool use and agent orchestration, as well as Segment Anything perception research associated with MSL FAIR. Muse Spark marked a consequential change in Meta’s model development: Meta described it as the first product of a ground-up overhaul of its AI efforts. In 2026, MSL released Muse Spark and Muse Spark 1.1, expanding developer access from a private API preview to a public preview of the Meta Model API.[8][9][3]

MSL introduced Muse Spark on April 8, 2026, making it available through the Meta AI app and meta.ai, with a private API preview for select users. The natively multimodal reasoning model supports tool use, visual chain of thought and multi-agent orchestration. Meta called it the first product of a “ground-up overhaul of our AI efforts,” following nine months of rebuilding the pre-training stack through changes to architecture, optimization and data curation.[8]

The overhaul addressed three distinct stages: pre-training, which establishes core capabilities; reinforcement learning, which improves them after pre-training; and test-time reasoning, which spends compute while answering. Muse Spark’s reinforcement learning maximizes correctness while penalizing thinking time to encourage more efficient reasoning. Its Contemplating mode uses multiple agents reasoning in parallel, providing another way to spend inference compute beyond extending a single agent’s reasoning.[8]

For health reasoning, Meta reported collaborating with more than 1,000 physicians to curate training data. It also announced investment across research, model training and infrastructure, including the Hyperion data center, to support further model scaling.[8]

On July 9, 2026, MSL released Muse Spark 1.1 in Thinking mode through the Meta AI app and meta.ai, and launched the Meta Model API in public preview. The release focused on agentic tasks, coding, computer use and multimodal understanding. MSL described a main agent that gathers context, plans and delegates execution to parallel subagents; subagents can escalate work back to the main agent.[9]

Muse Spark 1.1 has a one-million-token context window, with retrieval and compaction designed to retain critical steps from earlier work. Its computer-use training combines scripts, direct interface interaction and batches of actions, choosing between automation and clicking according to the task. MSL’s demonstrations included debugging a web application using screenshots and turning smartphone video into a Facebook Marketplace listing. The organization also reported daily use by Meta developers and researchers, including automation of model-development and evaluation tasks.[9]

The Muse family also expanded into media generation. MSL’s July 9 announcement confirmed that week’s Muse Image launch; Latent.Space’s July 8 recap reported the image-model launch alongside a preview of Muse Video.[9][4]

Before the Muse releases, Meta’s AI activity included investment in Scale AI and the established Segment Anything research program. Reuters reported on June 30, 2025 that Meta had invested $14.3 billion in Scale AI earlier that month. That figure was a parent-company investment in Scale AI, rather than financing raised by MSL. By the December 18, 2025 SAM interview, SAM lead Nikhila Ravi and SAM 3 researcher Pengchuan Zhang were identified under MSL FAIR. Segment Anything’s earlier generations therefore provide context for the perception research represented within MSL, alongside its later Muse model development.[10][3][8]

Segment Anything’s earlier generations supplied technical context for MSL FAIR’s perception work: SAM 1 used an 11-million-image data engine, while SAM 2 introduced memory-based video tracking. In a December 18, 2025 interview, Nikhila Ravi was identified as Meta’s SAM lead and Pengchuan Zhang as a SAM 3 researcher. They represented a larger collaborative team: the SAM 3 paper lists 38 authors, including Nicolas Carion, Ravi, Zhang and Christoph Feichtenhofer.[3][11]

SAM 3, first submitted on November 20, 2025, detects, segments and tracks objects using concept prompts: short noun phrases such as “yellow school bus,” image exemplars or both. It returns segmentation masks and unique identities for matching instances. An image-level detector and a memory-based video tracker share a backbone. A presence head separates recognition—whether an object is present—from localization—where it is—while the tracker preserves object identity through video.[11][3]

The paper’s March 28, 2026 revision describes a data engine producing four million unique concept labels, including hard negatives, across images and videos, and an open-source release of SAM 3 with the Segment Anything with Concepts benchmark. The December interview summary reported annotation stages falling from two minutes per image with human annotation to 45 seconds with model proposals and 25 seconds with AI verification fine-tuned on Llama 3.2. It also reported image inference at 30 milliseconds for 100 detected objects on an H200 GPU.[11][3]

Roboflow provided a downstream example of segmentation at scale. The December interview, featuring Roboflow CEO Joseph Nelson, reported 106 million smart polygons created at Roboflow and estimated savings of more than 130 years of labeling time. Applications included cancer research, underwater trash cleanup and autonomous vehicle perception; these figures describe partner usage and estimated labeling savings.[3]

Agent coordination also appeared in Strategy Auctions for Workload Efficiency (SALE), attributed to MSL research by Latent.Space’s February 6, 2026 recap. The paper, by Lisa Alazraki, William F. Shen, Yoram Bachrach and Akhil Mathur, was first submitted on February 2. Its June revision describes agents bidding with short strategic plans, scored through a cost-value mechanism and refined using shared auction memory. This routes each task without training a separate router or running every candidate model to completion.[5][12]

In March, the Dreamer team announced it was joining MSL. A March 23 update to David Singleton’s interview, published March 20, identified the recording as a snapshot before the transition. Singleton had left Stripe in 2024 and joined forces with Hugo Barra to build the startup initially known as /dev/agents. Latent.Space subsequently characterized the move as an “execuhire,” its term for licensing-plus-hiring arrangements.[7][1]

Before the transition, Dreamer was a consumer platform for discovering, building and using agents and applications. Its personal Sidekick planned builds, checked available tools, generated code and tested results. Dreamer hosted the applications and their databases and model access, while an SDK and CLI let developers edit code locally. Singleton described Sidekick as mediating interactions between agents according to the user’s permissions and expectations.[7]

Singleton demonstrated a conference planner built from the AI Engineer conference’s published data, using personal interests and Sidekick’s memory to select sessions. A separate ski-trip application combined live conditions, group information and expense settlement calculations; he explicitly said it did not move money. He reported that roughly six people had built the demonstrated platform and that Dreamer then had about 17 people. These were pre-transition Dreamer figures and demonstrations, rather than MSL headcount or announced Meta product integrations.[7]