Meta has unveiled Muse Spark, the first artificial intelligence model from its newly formed Superintelligence Labs team, created last year to close the gap with rivals in the AI race.
Be the Common Intelligence Layer Across Meta’s Ecosystem
Internally known under the codename “Avocado,” Muse Spark is the first model in a new family. It is designed to be fast, while still capable of reasoning through complex questions in science, mathematics, and health, with a solid foundation on which the next generation is already being built.
Meta is not disclosing its size, a key metric typically used to compare the computational heft of competing systems.
The model will power the Meta AI app and the meta.ai site, with deployment planned in the coming weeks across WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban Meta smart glasses.
The stated objective is to make Muse Spark the common intelligence layer for the entire group’s ecosystem, from social networks to connected devices.
Computational Efficiency, the Central Selling Point
Technically, the efficiency gain is one of the strongest selling points. Meta says Muse Spark achieves its reasoning capabilities using more than an order of magnitude less compute than Llama 4 Maverick.
This efficiency relies on a process called “thought compression.” During reinforcement learning, the model is penalized for excessive thinking time, forcing it to solve complex problems with fewer reasoning tokens without sacrificing accuracy. Multimodal and multi-agent, health applications are a priority axis. The clinical dataset underpinning these capabilities was developed with help from more than 1,000 doctors.
The model offers two operating modes: a fast mode for routine queries, and a reasoning mode for complex tasks.
A third, the “Contemplating” mode, will be rolled out gradually. It runs multiple agents in parallel to multiply the power of reasoning. An advanced illustrative use case highlighted by Meta: planning a family vacation with one agent drafting the itinerary while another searches for child-friendly activities.
Benchmarks: Top-5 Worldwide, but with Blind Spots
On the Artificial Analysis Intelligence Index, Muse Spark scores 52 (versus only 18 for Llama 4 Maverick), placing it just behind Gemini 3.1 Pro Preview (57), GPT-5.4 (57), and Claude Opus 4.6 (53).
Its strengths are 39.9% on Humanity’s Last Exam, 80.5% on MMMU-Pro for vision, 42.8 on HealthBench Hard, and 86.4 on CharXiv Reasoning.
The choice to pursue a proprietary model marks a break with the Llama models. Meta is testing a new revenue stream by offering third-party developers access to the technology via an API, currently limited to selected partners, with broader paid access to be announced later.
Shopping features are integrated directly into Meta AI, guiding users toward products to buy. With more than 3.5 billion active users across its platforms, Meta has enormous distribution leverage.
Alexandr Wang, the head of Superintelligence Labs, noted that more powerful versions are in development and that Meta plans to publish at least some of them as open source. A concession deemed necessary to a developer community that had been the principal differentiator of the Llama ecosystem.
