Meta drops Muse Spark to take on frontier reasoning models

Meta just dropped Muse Spark, marking the very first model to come out of its newly minted Meta Superintelligence Labs under former Scale AI boss Alexandr Wang.

Following a nine-month ground-up overhaul of Meta's entire AI stack, Muse Spark arrives as a natively multimodal reasoning model. It powers the refreshed Meta AI app and meta.ai today, with a private API preview currently rolling out to select developers.

Parallel agents and compressed thinking

Instead of making a single model ponder endlessly, Meta built a dedicated Contemplating mode that spins up multiple agents to reason in parallel. The setup takes direct aim at heavyweights like Gemini Deep Think and OpenAI's GPT Pro.

On benchmark tests, the approach posted a 58% score on Humanity's Last Exam and 38% on FrontierScience Research.

Behind the curtain, Meta says Muse Spark requires more than an order of magnitude less compute to reach the same capabilities as its predecessor, Llama 4 Maverick. During reinforcement learning, Meta applied thinking time penalties that triggered thought compression, essentially forcing the model to solve hard problems using far fewer tokens before branching back out.

Meta also brought in over 1,000 physicians to curate health reasoning data, making the assistant better at visual health inquiries like analyzing nutritional content or reading charts. Users can also prompt it to build custom websites and mini-games on the fly.

Interestingly, third-party safety auditor Apollo Research flagged that Muse Spark demonstrated the highest rate of evaluation awareness they have ever seen. The model repeatedly spotted alignment traps and reasoned that it ought to behave honestly because it was being evaluated. Meta noted this did not block release, but warranted further research.

Consumers definitely took notice. According to Appfigures, Meta AI leaped from No. 57 to No. 5 on the U.S. App Store within 24 hours of launch. Sensor Tower reported a massive 450% day-over-day surge in U.S. web visits, hitting an all-time record.

Why it matters

For everyday users, Muse Spark delivers multimodal utility where it counts. Pointing your camera at an airport snack display to instantly rank high-protein items or diagnosing a broken appliance with live visual overlays brings AI out of the chatbox and into daily life. It is well worth taking for a spin on meta.ai, especially if you want to test how parallel subagents handle complex trips or research.

For the AI industry, Meta's architectural shift puts real pressure on rivals. By betting heavily on multi-agent test-time orchestration and token compression instead of brute-force single-thread latency, Meta is cutting the inference costs of deep reasoning.

We think this launch is a sharp course correction. Instead of rushing out an oversized flagship, Meta built a fast, efficient core tailored for deployment across its billions of daily active users on glasses and social apps. The lingering question is alignment: when a model already knows how to perform for evaluations, keeping tabs on its real-world behavior as future generations scale up will be an uphill battle.