The model is the first major release since Alexandr Wang joined Meta nine months ago to lead AI efforts under the newly formed Meta Superintelligence Labs. Wang, former CEO of Scale AI, joined Meta as part of its $14.3 billion investment.
2 View gallery
Meta CEO Mark Zuckerberg during Meta AI presentation
(Photo: Reuters)
Dubbed Muse Spark and originally codenamed Avocado, this model is the first entry in Meta’s new Muse series. This comes after a disappointing rollout of the company’s previous Llama 4 model, which failed to gain support from developers and forced CEO Mark Zuckerberg to change strategy.
“Over the past nine months, Meta Superintelligence Labs has rebuilt our AI stack from the ground up, moving faster than any other development cycle we have ever run,” Meta said in a blog post. “This initial model is small and fast by design, but powerful enough to reason about complex questions in science, mathematics, and health. This is a strong foundation, and the next generation is already in development.”
Meta doesn’t position Muse Spark as a top-of-the-line model, instead focusing on efficiency and “competitive performance” across tasks such as inference, multimodal cognition, and health-related analytics.
Despite investing heavily in generative AI to support its advertising business and internal operations, Meta has struggled to establish a leading position in the AI model market. Rivals like OpenAI and Anthropic now have combined valuations of more than $1 trillion, while Google’s Gemini platform continues to expand in the consumer space.
The company has significantly increased spending to compete. Meta said it expects AI-related capital spending to reach $115 billion to $135 billion in 2026, nearly double last year’s total.
Muse Spark will be a proprietary model, but Meta said it may open source future versions. This marks a shift from the previous strategy using the Llama model, which was released as open source.
Meta said advances in training technology and infrastructure have made it possible to build smaller models with functionality comparable to previous medium-sized systems while using far less computing power.
“Muse Spark provides competitive performance in multimodal perception, inference, health, and agent tasks,” said Mehta. “We will continue to invest in areas where we currently have performance gaps, particularly in our long-term agent systems and coding workflows.”
The company is also exploring new revenue streams by offering developers access to Muse Spark through an application programming interface. Currently, private previews are only available to select partners, but Meta plans to expand paid access more broadly.
The model powers the AI assistant across Meta’s standalone apps and websites, with broader integration planned across Facebook, Instagram, WhatsApp, Messenger, and the company’s Ray-Ban smart glasses in the coming weeks.
Mehta said the updated assistant allows users to switch between different modes depending on the complexity of the task, from quick responses to more advanced analysis of things like legal documents and product information.
According to the company, “contemplation mode” will be gradually rolled out for more complex queries, using multiple AI agents for “parallel inference” and “competing with extreme inference modes in frontier models such as Gemini Deep Think and GPT Pro.”
The system will also include a shopping feature designed to help users discover and purchase products within Meta’s platform.
“Shopping mode draws from the styling inspiration and brand storytelling already happening across our apps, and surfaces ideas from the creators and communities people already follow,” Mehta said.
