Meta has unveiled its first new AI model in a year, the Muse Spark, which is the first in a new series of models internally called Avocado. The company is trying to get back into the AI race after the disappointment of Llama 4. Independent tests show that the model is not inferior to competitors in some areas, but still lags behind in coding and logical thinking.
Meta has unveiled its first new AI model in a year, the Muse Spark, which is the first in a new series of models internally called Avocado. The company is trying to get back into the AI race after the disappointment of Llama 4. Independent tests show that the model is not inferior to competitors in some areas, but still lags behind in coding and logical thinking.
American tech giants are now under pressure to prove that their huge spending on AI will pay off.
The stakes are particularly high for Meta, which hired Scale AI CEO Alex Wang last year in a $14.3 billion deal and offered individual engineers compensation packages in the hundreds of millions of dollars, putting Meta in charge of building a new superintelligence team as it seeks to return to the forefront of the industry after the disappointing results of its Llama 4 models earlier last year.
Muse Spark is the company's first release in a year. Initially, the model will be available only in the little-known Meta AI app and website. In the coming weeks, it will replace the current Llama models, which power chatbots on WhatsApp, Instagram, Facebook, and Meta's line of smart glasses, Reuters reports .
Meta did not disclose the size of Muse Spark, a key metric commonly used to compare the computing power of AI systems with competitors. The company also changed its open access strategy : while previous Llama models were open, Muse Spark only provided “private access” to a limited number of unnamed partners.
"This first model is designed to be small and fast, yet powerful enough to solve complex problems in science, math, and healthcare. It's a solid foundation, and the next generation is already in development," the company said in a blog post.
Independent performance evaluations of Muse Spark have shown that in some areas, such as language understanding and visual content, it is catching up with top models from market leaders Google, OpenAI, and Anthropic. However, the model still lags behind in other areas, such as coding and abstract reasoning.
In the overall AI test index compiled by the analytical firm Artificial Analysis, the model tied for fourth place with other competitors.
Wang acknowledged that “there are definitely some rough edges in the model’s behavior that we will polish over time.” He added that more powerful versions of the model are already in development, and Meta plans to release at least some of them into the public domain.
Meta plans to release at least some of these to the public. With this release, the company gave a clearer understanding of how it plans to monetize its models: Meta announced shopping features integrated into the Meta AI chatbot that will direct users directly to purchasable products.
Overall, the company is betting that applying AI to everyday personal tasks will increase engagement among its more than 3.5 billion users on its social platforms, potentially giving Meta an advantage over competitors with smaller audience reach.
The company also said that Muse Spark can help users with tasks such as counting calories in a dish based on a photo or overlaying an image of a mug on a shelf to see how it would look in an interior.
An additional Contemplating Mode, in which multiple agents work simultaneously to enhance analytical capabilities, will allow Muse Spark to compete with deep thinking modes such as Google's Gemini Deep Think and OpenAI's GPT Pro.
Meta noted that this mode can be used, for example, for effective family vacation planning: while one agent is creating a travel itinerary, another will simultaneously search for leisure options suitable for children.
