Meta launched Muse Spark on April 8 as its first fully closed AI model, reversing the open-source strategy that defined the Llama era. The release is the first product from Meta Superintelligence Labs, the unit formed around Alexandr Wang after Meta’s $14.3 billion investment in Scale AI.
Wang said the team rebuilt Meta’s AI stack from scratch over the past nine months, including infrastructure, architecture, and data pipelines. He added that larger models are already in development and that Meta still intends to open-source future versions, but Muse Spark itself is being released without public weights.
A closed debut replaces the old Llama playbook
Unlike Llama, Muse Spark does not make its weights publicly available. API access is also restricted for now, with invitations limited to selected partners. Meta has framed that decision as temporary, yet the launch still marks a sharp break from the company’s prior positioning around open models. Gartner analyst Arun Chandrasekaran called it a “major shift” and said the move suggests Meta may be moving away from the Llama brand altogether.
The reversal stands out because Llama had already reached 1.2 billion downloads by early 2026. Even so, Meta chose to release a proprietary model first, showing a clear change in how it is balancing ecosystem reach against product control and revenue capture.
Multimodal design with a health reasoning focus
Muse Spark is natively multimodal, handling text, image, and voice inputs. Its headline feature is a “Contemplating” mode that runs multiple reasoning agents in parallel before returning an answer, putting it in direct competition with Gemini Deep Think and GPT Pro.
Meta said it worked with more than 1,000 physicians to curate health-related training data, and the company is marketing the model as a personal health reasoning tool as well as a general assistant. On the Artificial Analysis Intelligence Index, Muse Spark scored 52, below the 57 posted by GPT-5.4 and Gemini 3.1 Pro. Meta has not disclosed parameter count or architecture details. According to the report, Muse Spark outperformed Gemini 3.1 Pro on several health-focused benchmarks that Meta prioritized in its own evaluation suite.
Competitive pressure is reshaping Meta’s AI strategy
Meta had already hinted at a phased rollout for its next generation of AI systems, keeping some core components proprietary while reviewing safety risks. A fully closed first release makes that direction concrete. It also reflects the pressure coming from OpenAI and Anthropic, whose proprietary models are generating billions of dollars in API revenue that Meta’s earlier open-source approach could not monetize in the same way.
Spending plans point to the same shift. Meta guided 2026 capital expenditure to $115 billion to $135 billion, nearly double 2025 levels. On the day Muse Spark launched, Meta shares rose more than 9%, the company’s strongest single-day market response to a product announcement in over two years. Developers who built on Llama are now being asked to wait for a future open-source release, and Meta has not provided a timeline.
Meta said Muse Spark will be rolled out across WhatsApp, Instagram, Facebook, and Messenger in the coming weeks.