Meta has announced the launch of Muse Spark, a new large language model built in house for the social media firm’s products. The company described it as the first in a new line of Muse large language models developed by its Meta Superintelligence Labs division, and said it will act as a small-scale base for future models.
The model now powers the Meta Artificial Intelligence assistant in its own app and website. Meta said it can handle complex problems in “Thinking mode” by running multiple subagents in parallel to solve discrete parts of a long prompt at the same time, speeding up responses. Muse Spark can also respond to visual prompts through what Meta calls multimodal perception. Meta said the model can rank a shelf of snacks by protein content or estimate the calories in a plate of food from a photo, and added that the feature was developed in collaboration with a team of physicians to support detailed answers to health questions.
Muse Spark can also build websites and mini-games from simple prompts. Meta is entering a field where similar capabilities are already standard in flagship models from Anthropic, OpenAI, and Mistral, but its own model is more tightly connected to the company’s social platforms. Meta said its shopping mode can source products based on “the styling inspiration and brand storytelling” taking place on Instagram, Facebook and Threads. The model can also aggregate posts to sharpen recommendations, including finding coffee shops in an area with large amounts of positive posts.
The upgraded modes, Thinking and Instant, are currently available on the Meta Artificial Intelligence app and meta.ai website, while the new features are in the process of being rolled out in the US. Meta said that it will bring the features to more countries and services over the coming weeks, including WhatsApp, Instagram and its Artificial Intelligence glasses.