Meta on Wednesday unveiled a new flagship artificial intelligence model called Muse Spark, its first under the revamped A.I. division that Mark Zuckerberg spent billions to build last year as the company pushes to catch up to rivals.
On tests that measure writing and reasoning, Muse Spark performed significantly better than Meta’s previous A.I. models and was nearly as good as the top models from competitors like Google, OpenAI and Anthropic, according to data provided by Meta.
But the model still lagged on coding ability, which has become the focus of Anthropic and others in the wider A.I. race. Meta’s announcement comes one day after Anthropic said its latest A.I. model, Mythos, was too powerful to safely release because of cybersecurity threats.
How Meta’s new A.I. model performs is being closely watched. Anthropic, Google and OpenAI are regarded as the leaders in developing foundational A.I. models and Mr. Zuckerberg, Meta’s chief executive, has pledged to catch up at any cost.
Last year, he shelled out billions to recruit a group of A.I. specialists to form a “superintelligence” lab to help with the effort. The model is a test of that new leadership under Alexandr Wang, the 29-year-old Silicon Valley entrepreneur who Mr. Zuckerberg hired to be the chief A.I. officer.
Mr. Zuckerberg, 41, has said that he will also invest $600 billion on new data centers to win the A.I. race. This year, the company — which owns Facebook, Instagram and WhatsApp — has forecast that it will spend as much as $135 billion, nearly twice the $72 billion it spent last year, much of it on A.I. Being at the forefront of A.I. development helps companies recruit talent and keep up a stream of experimentation.
“The new model and how it performs is really at the center of Meta’s A.I. credibility,” said Mike Proulx, a research director at Forrester VP, a research firm. “It’s the first real test of whether its massive A.I. investment can translate into a model that can stand alongside competition.”
Muse Spark, which had been referred to internally as Avocado, is now available on Meta’s stand alone A.I. app and will be coming to WhatsApp, Instagram and its A.I. smart glasses in the coming weeks, the company said.
While Meta has long open sourced its A.I. models, which means the underlying code is made public for developers to build on, Muse Spark is closed, which means the code is kept private. Meta said it may open source parts of the model in the future.
Mr. Zuckerberg revamped Meta’s A.I. division after its last A.I. model, Llama 4, fell short of expectations when it was released a year ago. Executives in charge of past models have since left the company, replaced by Mr. Wang and other new hires from companies like OpenAI.
Last summer, Mr. Zuckerberg declared that Meta’s new goal was to create a “superintelligent” form of A.I., a godlike technology that can act as the ultimate personal companion. Amid that push, Meta has laid off hundreds of employees and scaled down its investment in areas such as the metaverse and virtual reality.
Internally, Muse Spark was a nine-month sprint that included tensions and delays. In recent months, Mr. Zuckerberg tempered expectations for the new model, saying in January that it would “show the rapid trajectory we’re on” but not yet push the frontier of development.
Meta echoed that sentiment on Wednesday, saying it was still working to improve coding skills but had larger, more powerful models in the works. Its next model, which is in development, is known internally as Watermelon.
“This initial model is small and fast by design, yet capable enough to reason through complex questions in science, math and health,” Meta said. “It is a powerful foundation, and the next generation is already in development.”
- Credits: The New York Times
- Author: Eli Tan
- Photo: Adrienne Grunwald
