Meta has officially launched Muse Spark, its latest AI model and the first major product to emerge from its Superintelligence Labs. CEO Mark Zuckerberg personally announced the release, describing it as the opening move in a complete overhaul of the company's AI strategy. The model is now live on the Meta AI app and website.
Built for Everyday Tasks
Muse Spark is designed with practical, consumer-facing use cases in mind. It handles health-related queries, shopping assistance, visual understanding, and social content interactions. A dedicated shopping mode combines AI with data on individual user behaviour and interests — a clear nod to Meta's advertising roots. The model accepts voice, text, and image inputs, though it currently produces text-only responses. A fast mode handles casual queries while multiple reasoning modes tackle more complex requests.
A Break from Llama
Muse Spark marks a deliberate departure from Meta's earlier Llama models, which had consistently trailed rivals like OpenAI and Anthropic on key benchmarks. Zuckerberg, reportedly frustrated with that progress, initiated a structural overhaul. He brought in Alexandr Wang, former CEO of Scale AI, to lead the new Superintelligence Labs, and invested $14.3 billion in Scale AI for a 49% stake. Meta also recruited over 50 researchers from OpenAI, Google, and Anthropic before reorganising its teams into smaller, focused units. The model itself, internally code-named Avocado, was built over roughly nine months under Wang's leadership.
Massive Financial Backing
The launch follows staggering levels of investment. Meta spent around $72 billion on AI in 2025, with projections suggesting that figure could climb to $135 billion in 2026. Despite this, questions remain over commercial returns. An MIT study found that most companies deploying AI have yet to see meaningful financial gains. Muse Spark is effectively Meta's answer to those concerns, its first real proof-of-concept after years of heavy spending.
Where It Stands Against Competitors
Meta released benchmarks comparing Muse Spark against models from OpenAI, Google, and Anthropic. The results were mixed. The model performs competitively on multimodal understanding and health information processing, but Meta openly acknowledges a gap in areas like coding. A "Contemplating" mode designed to improve reasoning through multiple coordinated AI agents has also been introduced, though it is not yet widely available.
Privacy and Open-Source Plans
To use Muse Spark, users must log in with a Facebook or Instagram account. Meta has not explicitly stated whether data from those accounts will feed into the AI, though the company's privacy policy places few restrictions on how shared data can be used, a concern worth noting as the model scales. On the other hand, Meta has confirmed plans to release an open-source version of Muse Spark, continuing its tradition of making select models publicly available to developers.
What's Next
Zuckerberg's long-term vision goes beyond a capable chatbot. He has spoken about building AI that acts as a "personal superintelligence" — systems that don't just answer questions but complete tasks on your behalf. Muse Spark is the first step toward that, with plans to expand the model across Facebook, Instagram, and WhatsApp. More advanced models in the Muse family are also in the pipeline.
Whether Muse Spark can close the gap with its rivals and justify Meta's enormous investment will be the defining question as the AI race intensifies.
