Meta Launches Muse Spark AI Model From New Superintelligence Division
Meta unveiled Muse Spark, its first AI model from the newly formed Superintelligence Labs, marking a shift from open-source to proprietary development.

Meta announced the release of Muse Spark, the first artificial intelligence model from its newly established Superintelligence Labs division. The model represents a significant departure from Meta's previous open-source approach with its Llama series, as Muse Spark is proprietary and available only through Meta's AI app and website.
The new model was developed under the leadership of Alexandr Wang, former Scale AI co-founder, who was recruited to head Meta's overhauled AI operations as Chief AI Officer. Meta CEO Mark Zuckerberg reportedly restructured the company's AI division in summer 2025 following mixed reception of the Llama 4 model.
Muse Spark features what Meta calls "visual chain of thought" capabilities, allowing it to process and reason through visual information alongside text. The model includes a "Contemplating" mode that uses multiple sub-agents to reason in parallel, competing with advanced reasoning models from Google and OpenAI. According to Meta's benchmarks, Muse Spark achieved a score of 52 on the Artificial Analysis Intelligence Index, compared to Llama 4 Maverick's score of 18.
The model will initially power Meta AI across the company's platforms, including WhatsApp, Instagram, Facebook, and Messenger. Meta plans to integrate the AI into specialized features such as Instagram shopping recommendations and health-related analysis capabilities developed in collaboration with over 1,000 physicians.
The transition to a proprietary model has drawn criticism from developers who relied on Meta's open-source Llama models. The Llama family had reached 1.2 billion downloads by early 2026 and was widely adopted for enterprise applications. Wang indicated on social media that Meta plans to open-source future versions of the Muse family, though no timeline was provided.
Meta claims Muse Spark achieves its performance using significantly less computational resources than previous models, employing a "thought compression" process during training. The model's efficiency and multimodal capabilities position it as Meta's bid to compete directly with leading AI systems from rivals including Google's Gemini and OpenAI's GPT models.