MENLO PARK: Meta Platforms on Wednesday unveiled Muse Spark, the first artificial intelligence model released by its Meta Superintelligence Labs team, marking the company’s first major public model launch in about a year. Meta said Muse Spark now powers the Meta AI assistant in the Meta AI app and on its website, with the model set to replace existing Llama systems used in WhatsApp, Instagram, Facebook, Messenger and the company’s AI glasses in the coming weeks.
The launch follows a broad overhaul of Meta’s AI organization last year, when the company created Meta Superintelligence Labs and elevated Scale AI chief executive Alexandr Wang to a central leadership role after Meta’s $14.3 billion investment in the startup. Earlier this year, Meta’s chief technology officer said the new lab had delivered its first key models internally. The company’s previous Llama 4 release had drawn criticism over performance as rivals introduced newer systems with stronger results in several benchmarks.
Meta described Muse Spark as a natively multimodal reasoning model designed to handle text and images, use tools and support multiple agents working at the same time. In company materials, Meta said the model was built to answer complex questions in science, math and health while remaining small and fast. The company also said Muse Spark underpins an upgraded Meta AI experience with separate Instant and Thinking modes, allowing users to shift between faster replies and more deliberate reasoning depending on the task.
Product Rollout Expands Across Meta Apps
Meta said the updated assistant can process photos and visual information more directly, enabling tasks such as estimating calories from an image of food, comparing products and answering questions about charts or other visual inputs. The company also introduced shopping features that draw on content across its platforms and said Meta AI can surface public posts related to places or trending topics. Meta said the new features are beginning to roll out in the United States on the app and website before expanding to other countries and services.
At the same time, Meta is limiting outside access to the underlying model. The company did not disclose Muse Spark’s size, a standard metric often used to compare AI systems, and said the model will initially be offered in private preview through an application programming interface to select partners. That marks a departure from Meta’s earlier pattern of broadly releasing Llama models. Meta said larger Muse models are in development and added that it hopes to open-source future versions.
Benchmark Results Show Mixed Performance
Independent evaluations released alongside the launch showed Muse Spark performing competitively with leading models in areas such as language and visual understanding, while trailing in coding and abstract reasoning. On a broad benchmark index compiled by Artificial Analysis, the model tied for fourth place, according to published results cited on the day of the announcement. Meta acknowledged that the first version has limitations and said additional work remains on model behavior, even as it positions Muse Spark as the foundation of a new model family.
The company also paired the release with an updated AI risk framework and said a preparedness report on Muse Spark would follow. Meta said the model was developed over roughly nine months as the company rebuilt its AI systems, data pipelines and product experience, and that the launch is the first public step in a broader Muse series. For users, the immediate change is that Meta AI across the company’s platforms will increasingly run on Muse Spark rather than Llama. – By Content Syndication Services.