Introducing Muse Spark: Scaling Towards Personal Superintelligence

Meta just dropped the first model in a new lineup called Muse, and the company is framing it as step one on a much longer road toward what it calls personal superintelligence. Muse Spark is live today on meta.ai and the Meta AI app, with a private API preview rolling out to select users. It's a natively multimodal reasoning model, meaning it was built from scratch to handle images and tools together rather than bolting vision onto a text model after the fact.

The headline feature is something Meta calls Contemplating mode, which runs multiple agents reasoning in parallel to squeeze out extra performance on tough problems. Meta says this lets Muse Spark go toe-to-toe with the extreme reasoning modes in rivals like Gemini Deep Think and GPT Pro. On Humanity's Last Exam, Contemplating mode hit 58%. On FrontierScience Research, it scored 38%. Meta is upfront that gaps remain in long-horizon agentic tasks and coding workflows, areas it says it's still investing in as bigger models get built.

What's arguably more interesting than the benchmarks is how Meta got here. The company spent the last nine months rebuilding its pretraining stack — architecture, optimization, data curation, all of it — and claims the payoff is stark: Muse Spark reaches the same capability level as Llama 4 Maverick using over ten times less compute. On the reinforcement learning side, Meta says its new stack avoids the instability that usually plagues large-scale RL, showing steady, predictable gains rather than erratic ones. There's also a neat wrinkle in how the model handles test-time reasoning: it initially gets smarter by thinking longer, then hits a phase where a length penalty forces it to compress its reasoning into far fewer tokens, and then it extends again to push performance higher.

On the applications front, Meta is leaning hard into everyday, personal use cases rather than pure lab benchmarks. The multimodal chops let Muse Spark do things like troubleshoot a broken appliance by annotating what it sees, or spin up quick minigames on the fly. Health is the other big push — Meta worked with over 1,000 physicians to curate training data so the model can explain things like the nutrition breakdown of a meal or which muscles a given exercise targets, with interactive visual displays rather than plain text answers.

Safety testing turned up something genuinely odd. Third-party evaluators at Apollo Research found that Muse Spark showed the highest rate of