TechCrunch: Meta AI app reaches No. 5 on App Store after Muse Spark launch

On April 8–9, Meta’s AI app jumped from position 57 to number 5 on the U.S. App Store, driven by the release of Muse Spark — the first model developed by Meta Superintelligence Labs, the team assembled around Alexandr Wang after his move from Scale AI. The app recorded around 46,000 U.S. iOS downloads in a single day, an increase of 87% day-over-day.

Muse Spark is a natively multimodal reasoning model with tool use, visual chain of thought, and multi-agent orchestration. Despite being small and fast by design, it achieves performance comparable to Llama 4 Maverick using more than ten times less compute. Practically, this means lower inference costs per request — a meaningful factor for any team thinking about unit economics when building AI-powered features at scale. The model supports visual coding (generating websites and mini-games from a text prompt) and powers new shopping features in Meta AI that surface styling and product recommendations within Instagram and Facebook. It is rolling out across WhatsApp, Messenger, and Meta’s AI glasses, with a private API preview available to select partners.

Why it matters for product managers

The download surge illustrates something worth noting: Meta went from a lagging position in consumer AI mindshare to a top-five App Store ranking within a day, without a conventional marketing campaign — the product announcement itself drove the movement. This is a signal about how model capability releases function as distribution events in the current market, not just technical milestones.

For teams building AI products, the Muse Spark launch reinforces two things. First, model efficiency is increasingly a product differentiator — the ability to match frontier performance at a fraction of the compute cost changes what is economically viable to offer users at scale. Second, Meta’s platform distribution across billions of users means Muse Spark will accumulate real-world usage data at a pace that smaller AI developers cannot replicate, which over time shapes where model quality improves fastest.