Meta unveils Muse Spark, first model from its superintelligence team
TL;DR: Meta has released Muse Spark, the first AI model produced by the superintelligence team it assembled last year in a $14.3 billion deal to hire Scale AI CEO Alex Wang. The model is “small and fast by design” and will replace Meta’s existing Llama models across WhatsApp, Instagram, and Facebook in coming weeks. Independent evaluations show it catching up with leading models in some areas while lagging in coding and abstract reasoning.
The release ends a roughly year-long gap since Meta’s last model launch, a period during which its Llama 4 models underperformed expectations and competitors pulled ahead. Muse Spark is the first in a series internally codenamed Avocado, built by a team that includes engineers offered pay packages worth hundreds of millions of dollars.
A cautious return
Meta has notably changed its release approach. Unlike previous open Llama releases, Muse Spark is available only as a “private preview” to unnamed partners, with broader availability initially limited to the Meta AI app and website. The shift away from open release is a significant departure for a company that had positioned open-source AI as a strategic differentiator.
Independent benchmarks from Artificial Analysis placed Muse Spark in a tie for fourth on a broad index of AI tests, showing strength in language and visual understanding but weakness in coding and abstract reasoning. CEO Mark Zuckerberg had set expectations carefully, telling investors in January that early models would be “good but, more importantly, will show the rapid trajectory that we’re on.”
The release also hints at Meta’s monetisation strategy: shopping features embedded within the Meta AI chatbot will point users directly to purchasable products, leveraging the company’s 3.5 billion-strong user base across its social platforms.
Looking forward
The competitive dynamics in frontier AI continue to intensify. With Google, OpenAI, and Anthropic all releasing major models this year, Meta’s re-entry adds a fourth well-funded competitor to the field. The trajectory of subsequent Avocado models will determine whether Meta’s massive investment in its superintelligence team delivers on its promise.