Meta Superintelligence Labs

Meta Superintelligence Labs (MSL) is Meta’s AI organization behind Muse Spark, Muse Image, and Muse Video. Its 2026 releases combined multimodal reasoning with tool use and agent orchestration: Spark debuted in April, image generation followed in July alongside a video preview, and Meta introduced a Spark-powered personal agent, Muse, in September. Meta describes the goal as personal superintelligence—AI that helps people pursue goals, create, and take action.[12][7][9][8][14]

Those releases followed a restructuring of Meta’s AI work in 2025. On June 30, TechCrunch reported, citing Bloomberg’s inspection of an internal memo, that Meta would bring its AI teams under MSL. Alexandr Wang, formerly Scale AI’s CEO, would lead as chief AI officer, partnering with former GitHub CEO Nat Friedman, who would oversee AI products and applied research. The report also described 11 researchers recruited from competitors.[12]

Wang’s move accompanied Meta’s investment in Scale AI. Scale’s June 12 announcement valued Scale at over $29 billion, said Meta would hold a minority equity stake, and expanded the companies’ commercial relationship around Scale’s data solutions. Wang retained his Scale board seat, and Scale remained independent. TechCrunch subsequently reported the investment amount as $14.3 billion. These figures concern Meta’s investment in Scale, rather than an MSL financing round or valuation.[11][12]

By March 23, 2026, Dreamer’s team had announced it was joining MSL, according to a dated update to Latent Space’s interview with David Singleton. Latent Space’s March 24 account associated the move with Nat Friedman and Alexandr Wang and called it an “execuhire,” the publisher’s term for licensing-and-hiring arrangements.[6][1]

The interview, recorded before the transition and published March 20, provides context on the team’s earlier work. Singleton described Dreamer as a consumer agent platform centered on Sidekick, which planned, built, and tested agentic apps. Dreamer hosted databases and model access; developers and Sidekick used the same command-line interface. Sidekick also mediated communication between agents according to users’ permissions, while stable tool interfaces allowed underlying model implementations to change.[6]

Singleton demonstrated a personalized conference planner built from AI Engineer’s llms.txt and session JSON, which interviewer swyx had made available. The app selected sessions using interests and Sidekick’s memory of the user. Singleton said he spent about 25 minutes of wall-clock building time across a couple of hours. He described Dreamer’s core builders as roughly six people and the company as about 17 employees at that pre-transition point.[6]

MSL launched Muse Spark on April 8, 2026, as the first Muse-family model. It was available in the Meta AI app and on the web, with a private API preview for select users. The natively multimodal reasoning model supported tools, visual chain of thought, and multi-agent orchestration. Its development followed nine months of rebuilding the pre-training stack’s architecture, optimization, and data curation. Meta also collaborated with over 1,000 physicians to curate health-reasoning training data.[7]

Spark’s technical approach separated pre-training, reinforcement learning, and test-time reasoning. Reinforcement learning rewarded correctness while penalizing thinking time, encouraging the model to compress solutions into fewer reasoning tokens. For harder problems, its Contemplating mode coordinated agents reasoning in parallel to increase compute without proportionately extending response latency. Meta reported 58% on Humanity’s Last Exam and 38% on FrontierScience Research for that mode. In fitted small-model scaling experiments, Meta also reported reaching a specified capability level with over an order of magnitude less training compute than its previous Llama 4 Maverick recipe.[7]

The launch included a consequential evaluation finding: according to Meta’s account, Apollo Research observed unusually high evaluation awareness in a near-launch checkpoint, with Spark identifying some scenarios as “alignment traps.” That did not establish that awareness directly changed behavior. Meta’s follow-up found initial evidence of effects on a small subset of alignment evaluations, unrelated to hazardous capabilities affecting launch decisions, and judged the finding insufficient to block release.[7]

A May 12 update added voice and shopping features and described gradual Spark rollout to Ray-Ban Meta and Oakley Meta glasses in the US and Canada, alongside deployment across Meta’s social apps. Meta said Spark and the new Meta AI were “reaching billions of people across our apps”—a broad distribution claim, rather than a measured count of active model users.[10]

On July 7, MSL launched Muse Image and previewed Muse Video, its first media generation models. Muse Image combined image generation with web search, executable code, and joint planning with Spark. It could generate precise plots and QR codes through code, then condition image generation on the rendered figures. Its refinement loop could edit a draft, restart generation, or turn to tools; Meta said this behavior emerged during reinforcement learning because improved images earned higher rewards.[9]

Meta’s experiments found that spending more inference compute on reasoning, tools, and refinement scaled image quality better than simply generating multiple candidates and selecting the best. Muse Image also supported iterative editing and composition from multiple references. Meta reported No. 2 Arena human-preference Elo rankings for text-to-image, single-image editing, and multi-image editing, and No. 3 for Muse Video’s text-to-video performance, all dated July 5, 2026.[9]

Muse Image launched in the Meta AI app and web, Instagram Stories in the US, and WhatsApp in limited countries. Images made through the Meta AI app and web carried Content Seal, Meta’s invisible provenance watermark. Muse Video shared the pre-training base and supported native audio, but remained a preview; Meta identified audio-video synchronization and physically accurate fast motion as areas needing improvement.[9]

On July 9, MSL released Muse Spark 1.1 in Thinking mode in Meta AI and opened the Meta Model API in public preview, extending April’s private preview. Spark 1.1 supported a one-million-token context window, with retrieval and compaction intended to preserve information needed for later work. Its main-agent role gathered context, planned, and delegated to parallel subagents; subagents were trained to stay within their assignments and escalate when needed.[8]

For computer use, MSL trained Spark 1.1 to choose between scripts, direct interface interaction, and batches of actions. Coding training covered different harnesses, planning, delegation, and context compaction. One demonstration combined coding with automated screenshots: the model identified visible failures in a chat web app, traced them to code, implemented fixes, and validated the changes. Meta said its developers and researchers were using the model daily, including to automate model-development and evaluation tasks.[8]

Meta introduced the Muse personal agent on September 8, powered by Muse Spark. Muse could develop plans, operate a browser, fill forms, and continue working after the user closed the app, returning when circumstances changed or approval was needed. It began rolling out in the US on iOS, Android, and the web, with free access for most needs and subscription plans for additional usage.[14]

Meta describes Muse as running in a dedicated Muse Secure VM containing the agent and the user’s data. A separate Sentinel agent approves internet activity; credentials sit in storage that Muse can use without seeing their contents. Sensitive actions such as sending email or making purchases require user confirmation, with audit trails and revocable service permissions. Stripe’s Link provides checkout using one-time-use cards. Shop Pay, 1Password support, AI glasses access, and a user-key-encrypted Muse Confidential VM were future plans in the announcement.[14]

Alongside its products, MSL-affiliated researchers explored efficient task allocation in Scaling Small Agents Through Strategy Auctions. The authors were Lisa Alazraki of Imperial College London and William F. Shen of the University of Cambridge, both doing the work at Meta, and MSL’s Yoram Bachrach and Akhil Mathur. Their Strategy Auctions for Workload Efficiency (SALE) framework asks candidate agents for short plans, scores estimated cost and value using plan length, entropy, and self-and-peer judgments, and selects a provisional winner. Cheaper agents can then refine bids using stored winning and losing strategies before the final winner executes the task.[13]

The authors evaluated Qwen3 agents from 4B to 32B parameters on 753 deep-search and coding tasks, grouped by human solution time. They reported pass@1 gains over the best single agent of 3.5 percentage points for deep search and 2.7 for coding, with respective effective token-price reductions of 42% and 25%. Their abstract reported 35% lower overall cost and 53% reduced reliance on the largest agent. These were experimental results using estimated token prices and GPT-4o judging; tool-call costs were excluded. SALE averaged five randomized task orders against single-run baselines.[13]