Say goodbye to Llama! Meta launches a brand-new multimodal AI model, “Muse Spark,” rolling out across IG, FB, and smart glasses

Meta’s Superintelligence Labs (Meta Superintelligence Labs) has officially released its new multimodal AI model, “Muse Spark.” This marks Meta’s triumphant return to the AI arms race with a major model built specifically for its own products, following the reshuffling of its AI team.
(Background: IG launched a paid “invisible peek at your ex’s Stories” subscription! Meta has started testing subscriber accounts)
(Additional context: Meta smart glasses face a federal class-action lawsuit—after allegations of privacy leaks, users have sued in court; what do lawyers think?)

Table of contents

Toggle

• Built specifically for the Meta ecosystem, supports multimodal reasoning
• Not open-sourced yet! Shifting to invited API testing
• A key step to restart the AI battlefront

After experiencing technical bottlenecks and a latency controversy from the Llama series, social media giant Meta has officially shown the world its first set of results after spending tens of billions of dollars to reorganize its AI team. According to an announcement released by Meta on April 8, its newly appointed leader, Alexandr Wang, and the “Meta Superintelligence Labs (Meta Superintelligence Labs)” have officially launched a new large language model—Muse Spark.

Built specifically for the Meta ecosystem, supports multimodal reasoning

Unlike the past Llama series strategy that emphasized fully open-sourcing, Muse Spark is currently strongly positioned as a “bottom-layer engine tailored specifically for Meta products.” In the official announcement, Meta states that Muse Spark is a model with strong multimodal (supports text and image processing) reasoning capabilities, aiming to significantly reduce the performance gap with competitors such as OpenAI, Anthropic, and Google.

At present, Muse Spark has already launched first in the Meta AI app in the U.S. region and on a dedicated website. Meta has previewed that it will fully and deeply integrate this new model into its “all-in-one” suite of apps for nearly 4 billion monthly active users, including WhatsApp, Instagram, Facebook, and Messenger. At the same time, this multimodal model will also serve as the core “brain” for Meta smart glasses (Smart Glasses), with the goal of greatly enhancing visual and voice interaction experiences for wearable hardware.

Not open-sourced yet! Switching to invited API testing

Worth noting is that this time Meta has not chosen to directly release the open-source weights of Muse Spark to the public as it did previously. Based on the announcement and current industry hands-on testing, Meta is only making a private API preview (Private API Preview) available to “selected partners.”

This shift in strategy has sparked lively discussion on platforms such as the developer community and Hacker News. Many insiders believe Meta may be re-evaluating its open-source baseline as it moves into the Superintelligence era. At the same time, it also shows how eager Meta is to monetize powerful AI models directly, or to internalize them as a moat for its community platforms.

A key step to restart the AI battlefront

Over the past nine months, Meta has poured massive computing power and resources into developing Muse Spark. With the release of this new model, Meta not only reiterates that it has not been absent from the AI battlefield, but also sets the direction for future development. In its article, the company emphasizes that Muse Spark is only the first step in the Meta Superintelligence Labs series of models. This means Meta’s AI strategy has officially upgraded from its early “open-source community-oriented” approach to a “product-first” commercialization and ecosystem-binding model.