Welcome back to the frontier AI model race, Meta. After spending billions of dollars (and churning significant staff) to overhaul its efforts in the space, Mark Zuckerberg’s company finally has a new model to show the world. Dubbed Muse Spark, the model represents a significant leap over the company’s previous underwhelming efforts and puts it back in the mix with competitors like OpenAI, Anthropic, and Google on most major benchmarks for model performance—though it’s not yet challenging for the top spot in most categories.
The model, built from scratch by Meta Superintelligence Labs under the leadership of Alexandr Wang, represents something of a fresh start for the company. It claimed the Muse Spark is a “natively multimodal reasoning model with support for tool-use, visual chain of thought, and multi-agent orchestration,” and is built to be integrated directly into other Meta’s products. As such, it’ll reportedly roll out across Facebook, Instagram, Messenger, and WhatsApp in the coming weeks.
What people can expect, if benchmarks are to be believed (Meta has been accused of fudging its benchmarks in the past and didn’t release a paper along with the model, so maybe take these figures with a heaping helping of salt), is a model that is significantly more capable than prior offerings across basically all categories. By Meta’s own marks, it ranks behind only Google Gemini 3.1 Pro and OpenAI’s GPT-5.4 in multimodal functionality, or the ability to process information across different data streams and formats. It was also competitive in reasoning tests, though it typically fell short of the high marks from Anthropic’s Claude, and the top Gemini and GPT models.
Where Muse Spark continues to struggle is with coding and agentic functionality, which allows the model to complete tasks autonomously. It’s functional in those areas in ways that LLaMa 4 wasn’t, but it’s still a ways off from putting any dent in Anthropic’s dominance in the vibe coding sphere.
Meta has two things it is hoping will allow its model to stand out. The first is shopping. Meta claims that Muse Spark is capable of drawing styling inspiration from creators and communities that people follow to offer personalized product recommendations. Affiliate-style sales have been one of the low-hanging fruits for monetization for AI firms, so it’s no surprise Meta is reaching for it.
The other area the company is emphasizing is its ability to process health data. Again, it’s a good example of Meta trying to score some easy points, as asking health questions is a common AI use case for consumers. But Meta might have a bit too much baggage when it comes to how the company collects and uses data from its users to get the kind of buy-in it’d need to become anyone’s pocket-bound physician.
The release of Muse Spark isn’t exactly earth-shattering—it’s an interesting choice for Meta to drop a perfectly capable model that doesn’t fundamentally shift the landscape while Anthropic is in the middle of yet another “Our model is so powerful it might destroy the world as we know it” campaign—but it at least gets Meta back into the mix. It also gives the company a higher baseline in most major categories that it’ll no doubt aim to build on with future releases. But for now, Zuckerberg’s multi-billion-dollar spending spree has moved Meta from out of the race entirely to competing for also-ran status.
