Meta Launches First AI Model From Restructured Superintelligence Unit

Muse Spark powers Meta AI across apps and glasses; closed-source model marks strategic shift from open-source roots

Meta Platforms on Tuesday released Muse Spark, the first large language model from Meta Superintelligence Labs, and deployed the system across its Meta AI assistant, flagging a broader rollout to WhatsApp, Instagram, Facebook, Messenger, and AI glasses in the coming weeks. 

The model is the first product to emerge from a nine-month overhaul of Meta’s AI development stack, the company said in a blog post. Meta Superintelligence Labs, the unit that built Muse Spark, is led by Chief AI Officer Alexandr Wang. 

Muse Spark is a closed model; its architecture and code will not be made public, a departure from Meta’s prior strategy of releasing open-source AI systems, including the Llama family. The company said it plans to offer API access through a private preview to select partners and hopes to open-source future versions.

The move comes as Meta attempts to close a perceived capability gap with rivals OpenAI and Google following a multibillion-dollar restructuring of its AI operations.

ALSO READ: Anthropic Appoints Amlan Mohanty to Head India Policy Operations

Muse Spark is built for multimodal tasks; the system can process images alongside text, enabling Meta AI to analyze photographs and generate context-aware responses. The company cited use cases including identifying nutritional content from photos of food and annotating health-related imagery. 

The model also supports multi-agent orchestration. Meta said users can trigger parallel subagents within a single query, for instance, one agent drafting a travel itinerary while others research destinations and activities simultaneously. 

A separate capability called Contemplating mode, which coordinates multiple reasoning agents in parallel, will roll out gradually at meta.ai. Meta said the model achieved 58% on Humanity’s Last Exam and 38% on FrontierScience Research benchmarks, positioning it against what Meta described as the “extreme reasoning modes” of competing frontier models. 

Meta said it collaborated with more than 1,000 physicians to develop Muse Spark’s health reasoning capabilities, enabling the model to respond to queries involving medical images and charts. Health queries, the company said, are among the top reasons users turn to Meta AI. 

Separately, a new shopping mode within Meta AI draws on content shared across Instagram, Facebook, and Threads to surface product and style recommendations tied to creators and communities users already follow. 

Meta said its new pretraining stack requires more than an order of magnitude less compute than its prior Llama 4 Maverick model to achieve comparable performance, a claim the company attributed to improvements in model architecture, optimization, and data curation. 

The Muse series is designed as a generational scaling ladder, with each model intended to validate the approach before larger versions are released. Meta said larger models are already in development.

Meta disclosed that third-party evaluator Apollo Research found Muse Spark exhibited the highest rate of what it called “evaluation awareness”  behavior, in which the model identifies that it is being tested, of any model Apollo has assessed. Meta said its own follow-up investigation found initial evidence that the behavior may affect a subset of alignment evaluations, though none involved hazardous capabilities. The company said the finding was not a release-blocking concern but warrants further research. Full results are expected in an upcoming Safety and Preparedness Report. 

The upgraded Meta AI app and meta.ai are available now in the United States. A broader geographic rollout and integration into Meta’s family of apps and AI glasses are expected in the coming weeks.

Disclaimer: This report is based on publicly available reports, including official blog posts from Meta. Nervnow has not independently verified the claims.

MORE ON META AI
Nebius-Meta $27B Deal Highlights Big Push in AI Infrastructure Expansion
Meta Acquires Moltbook, the AI-Only Social Network
Nasdaq Rebounds as Meta Teams With AMD
Meta, Nvidia Forge Landmark Multi-Year AI Infrastructure Deal