In a major step forward in artificial intelligence, Meta has introduced Muse Spark, the first model developed by its newly formed Meta Superintelligence Labs (MSL).
The initiative, spearheaded by Alexandr Wang under the leadership of Mark Zuckerberg, marks a significant shift in the company's AI strategy.
MSL was established in 2025 following Meta's multi-billion-dollar investment, including its acquisition of Scale AI. In less than a year, the team has delivered Muse Spark, a model Wang claims was built entirely from scratch in just nine months. Despite being described as a "small and fast" model, Meta positions it as its most advanced AI system to date.
Muse Spark belongs to a new generation of models under the "Muse" family, which are expected to eventually replace Meta's earlier Llama models. While it is not the flagship offering in the lineup, Spark is designed to handle complex reasoning tasks across domains such as science, mathematics, and healthcare.
According to Meta, the model demonstrates competitive performance against leading AI systems, including Anthropic's Claude Opus 4.6 and OpenAI's GPT 5.4, particularly in multimodal reasoning and agent-based workflows. Wang highlighted a unique feature called "contemplating mode," which allows multiple AI agents to operate simultaneously to solve intricate problems. He noted, "In our testing we found it competitive w/ other extreme reasoning models such as Gemini Deep Think & GPT Pro."
One of the defining aspects of Muse Spark is its integration across Meta's vast ecosystem. The model is set to power Meta AI experiences across platforms like Instagram, WhatsApp, Facebook, and Messenger, as well as its standalone app and website. The company is also exploring its use in smart glasses and introducing features such as a shopping mode that delivers personalized recommendations based on creators and brands.
Wang emphasized the model's real-world awareness, stating that Muse Spark can "understand the world around you" and draw insights from user interactions across Meta platforms. He added that the system is designed with scale in mind, built for the "3 billion people already using [Meta] apps every day."
However, unlike its Llama predecessors, Muse Spark is not open source. Meta has opted for a closed approach for now, offering the model through a limited private API preview to selected partners. The company has indicated that future versions may become open source, but no timeline has been confirmed.
Safety has also been a focus during development. Wang claimed that Muse Spark showed "strong refusal behaviour in high-risk areas like biological and chemical weapons" during testing, aligning with industry-wide efforts to ensure responsible AI deployment.
While this release represents only the first step for Meta's Superintelligence Labs, Wang confirmed that larger and more powerful models are already in development, supported by rapidly scaling infrastructure.