Meta has unveiled Muse Spark, a new proprietary artificial intelligence model developed by its recently established Meta Superintelligence Labs (MSL). This launch represents a significant departure from the company’s previous open-source approach, introducing a closed model designed to compete directly with industry leaders like OpenAI and Google. Following the announcement, the Meta AI platform reached an all-time high for daily web visitors in the United States. Furthermore, the Meta AI iOS app experienced an 87 per cent day-over-day surge with roughly 46,000 downloads, propelling it from an average rank of 65 to the number six spot among free applications on the App Store. This growth extended globally, with downloads soaring in Canada, the United Kingdom, France, and Germany.
Designed to enhance everyday personal tasks, Muse Spark is currently available on the Meta AI standalone app and website. In the coming weeks, the company plans to integrate the model across its massive ecosystem, including WhatsApp, Instagram, Facebook, Messenger, and its Ray-Ban smart glasses. This broad rollout aims to boost user engagement by applying advanced AI to daily activities.
A New Era of Multimodal Reasoning
Muse Spark is a natively multimodal reasoning model, meaning it can process and generate both text and images while working through complex problems step-by-step. Built from the ground up over nine months under the internal code name “Avocado,” the model features a completely overhauled technical stack. Meta claims these architectural and data curation improvements allow the model to achieve high capabilities with significantly less computing power compared to its predecessor, Llama 4 Maverick.
Users can interact with Muse Spark through three distinct reasoning tiers: Instant, Thinking, and Contemplating modes. The Contemplating mode allows the AI to launch multiple subagents that work in parallel to solve multi-part tasks. For example, when planning a family vacation, one subagent can draft the travel itinerary while another simultaneously researches kid-friendly activities.
The model’s visual capabilities allow users to upload images without needing to provide explanatory text. It can estimate the calorie count of a meal from a photograph or superimpose an image of a mug onto a physical shelf to visualise how it would look. Additionally, Muse Spark introduces visual coding features that can generate mini-games and custom websites based on simple text prompts.
Shopping and Recommendation Features
A major addition to the Meta AI experience is the new Shopping mode, which provides product ideas and styling inspiration . This feature curates recommendations directly from creators and communities already active on Meta’s social platforms, assisting users with tasks like buying clothes or decorating rooms .
Muse Spark also leverages community content to provide localized context during interactions . When users ask about a trending topic or a specific location, the AI can surface relevant public posts from locals who are familiar with the area, creating a more integrated social experience that connects AI assistance with user-generated content .
Performance Benchmarks and Safety Evaluations
Meta designed Muse Spark to be small and fast, yet highly capable in specific domains such as science, mathematics, and health. To improve medical accuracy, the company collaborated with a team of physicians during the development process.
According to benchmark tests published by Meta, the model performs competitively against rivals, though it does not dominate across all categories. On the Health Bench Hard evaluation, Muse Spark achieved a score of 42.8 per cent, outperforming Anthropic’s Claude Opus 4.6, Google’s Gemini 3.1 Pro, and OpenAI’s GPT-5.4. However, on the GPQA Diamond benchmark—which tests PhD-level reasoning—Muse Spark scored 89.5 per cent, trailing slightly behind those same frontier models. Meta executives have acknowledged these performance gaps, noting that the model is currently weaker at coding and long-horizon agentic workflows.
In terms of safety, Muse Spark underwent rigorous evaluations and successfully refused 98 percent of requests that could potentially aid in bioweapons engineering. Third-party evaluator Apollo Research observed that the model displayed a high rate of evaluation awareness, frequently identifying test scenarios as alignment traps, though Meta concluded this was not a barrier to public release.
Strategic Shift and Billion-Dollar Investments
The development of Muse Spark follows Meta’s $14 billion investment in Scale AI in June 2025, which brought Alexandr Wang to the company as its first Chief AI Officer. Wang now leads the Meta Superintelligence Labs, an autonomous unit with a streamlined management structure. In March 2026, Meta further reorganized by creating a new applied AI engineering division under Vice President Maher Saba to build a data engine that accelerates model improvements.
Unlike Meta’s previous Llama models, which were released as open-weight systems for public use and modification, Muse Spark is a closed, proprietary model. Its architecture and underlying code will not be made publicly available, reflecting Wang’s preference for closed systems. Interestingly, Meta utilized several third-party open-source models during Muse Spark’s training phase, including Alibaba’s Qwen, alongside models from OpenAI and Google.
While the Meta AI chatbot remains free, the company is reportedly exploring the possibility of introducing subscription fees in the future. For now, Muse Spark serves as the foundational step in a deliberate scaling strategy, with Meta confirming that its next-generation models are already in active development.
