After months of heavy spending, a controversial talent acquisition, and mounting pressure to prove it can compete with OpenAI, Anthropic, and Google, Meta has released its first major AI model under the new Meta Superintelligence Labs banner. The model, called Muse Spark, went live on Wednesday and is already powering the Meta AI app and website, with a wider rollout to WhatsApp, Instagram, Facebook, Messenger, and its Ray-Ban AI glasses expected in the coming weeks.
Nine Months to Build, Billions to Back It
Meta Superintelligence Labs was created after CEO Mark Zuckerberg grew frustrated with the company’s Llama models and how they lagged behind rivals. To fix that, Meta spent $14.3 billion in June 2025 to acquire a 49% non-voting stake in Scale AI and brought in its co-founder and CEO, Alexandr Wang, as Meta’s first-ever chief AI officer. Wang and Zuckerberg then went on an aggressive talent hunt, offering pay packages at competing AI labs that reportedly climbed into the hundreds of millions of dollars when equity was included.
The result is Muse Spark, the first model in Meta’s new Muse series, described as a deliberate, scientific approach to model scaling where each generation validates and builds on the last before the company goes bigger.
Small Model, Significant Claims
Meta is not positioning Muse Spark as a top-of-the-line product. The company describes it as small and fast by design, yet capable enough to reason through complex questions in science, math, and health. That efficiency is central to the pitch. Meta claims its advances allow it to achieve the same capabilities with over an order of magnitude less compute than Llama 4 Maverick, its previous model.
On the features side, the model leans into multimodality. Muse Spark can see and understand what users are looking at, for example, snapping a photo of a product shelf and getting a nutritional ranking without typing any labels. The model also introduces a Shopping mode, drawing from styling inspiration and brand storytelling across Meta’s platforms to surface personalised recommendations from creators and communities people already follow.
A Closed Door for Open-Source Fans
Perhaps the most striking shift with Muse Spark is what it is not. In a pivot from Meta’s prior open-source strategy, Muse Spark is a closed model, meaning its design and code will not be made public. Meta has said it hopes to open-source future versions, but for now, the model will only be available to select partners through a private API preview.
This is a meaningful break from the Llama era, when open-source availability was central to Meta’s identity in the AI space and a key differentiator from its rivals.
Competitive, But Not the Top Dog
According to benchmark tests Meta published, Muse Spark is competitive with leading models from OpenAI, Anthropic, and Google across many tasks, although it does not surpass them across the board. Independent verification of those claims remains pending, and there is reason for caution. Meta was previously caught manipulating benchmark results for its Llama 4 models, later admitting to using specialised, unreleased versions fine-tuned for specific tasks to boost scores.
The health features may also attract scrutiny. Users must log in with a Facebook or Instagram account to access Muse Spark, raising questions about how personal data from those accounts could be used to train or personalise the model.
Reclaiming Ground, Not Leading It
Muse Spark is a credible recovery from the Llama 4 stumble, and it signals that Meta’s expensive restructuring is beginning to produce tangible output. But recovery is not leadership. The model puts Meta back in the mix with major competitors on most benchmarks, though it is not yet challenging for the top spot in most categories. With capital expenditure for AI-related infrastructure projected at between $115 billion and $135 billion for 2026, the stakes for what comes next in the Muse series are extremely high.
