Meta Superintelligence Labs on April 8 introduced Muse Spark, a multimodal reasoning model with tool use, visual chain-of-thought, and multi-agent orchestration.
The company said it is the first release in its Muse model family built for “personal superintelligence.”
Muse Spark currently powers Meta’s AI app and website and will roll out to WhatsApp, Instagram, Facebook, Messenger, and AI glasses in the coming weeks.
The company said a private API preview will be made available to select users.
“Muse Spark is the first step on our scaling ladder and the first product of a ground-up overhaul of our AI efforts,” the company said, adding that it is investing across research, training, and infrastructure, including the Hyperion data centre.
The company said Muse Spark delivers performance across multimodal perception, reasoning, health, and agent-based tasks, while continuing work on long-horizon agentic systems and coding workflows.
A new “Contemplating mode,” which runs multiple agents in parallel, improves performance on complex tasks. Meta said the mode achieved 58% on Humanity’s Last Exam and 38% on FrontierScience Research benchmarks.
The system is positioned to compete with reasoning-focused models such as Gemini Deep Think and GPT Pro, according to the company.
Meta said Muse Spark supports applications including visual problem-solving, interactive content creation, and health-related analysis.
The company noted it worked with over 1,000 physicians to improve health reasoning capabilities, enabling outputs such as nutritional breakdowns and exercise-related insights.
The release also reflects changes in Meta’s model development approach, focusing on three scaling axes, including pretraining, reinforcement learning (RL), and test-time reasoning.
In pretraining, Meta said it rebuilt its stack over nine months, improving architecture, optimisation, and data curation. “We can reach the same capabilities with over an order of magnitude less compute than our previous model,” the company said, referring to Llama 4 Maverick.
For post-training, Meta highlighted RL as a method to “scalably amplify model capabilities,” reporting “smooth, predictable gains” in accuracy and reliability. It added that improvements generalised to tasks outside the training data.
At inference, the company said it optimises reasoning efficiency using “thinking time penalties” and multi-agent orchestration. This allows the model to reduce token usage while maintaining performance, and to scale reasoning through parallel agents without increasing latency.
On safety, Meta said Muse Spark was evaluated under its Advanced AI Scaling Framework, covering risk categories such as cybersecurity, biological threats, and loss of control. “Muse Spark demonstrates strong refusal behaviour across high-risk domains,” the company said, adding that the model did not show capabilities required to realise threat scenarios.
ALSO READ: The Playground is Closed: 10 Hard Truths from the Cisco AI Summit
