Meta, Facebook's parent company, unveiled its next-generation artificial intelligence (AI) model, "Muse Spark." It is the first model introduced after Meta CEO Mark Zuckerberg carried out a sweeping AI reorganization to catch up with rivals such as OpenAI and Google.
Meta said on the 8th (local time) that it released "Muse Spark," the first AI model developed by the Meta Superintelligence Lab (MSL), which it established last year. Meta said, "Muse Spark is the most powerful model Meta has introduced to date," adding, "It is designed to be small and fast, yet it can reason through complex questions in science, math, and health."
Unlike Meta's previous open-source (developer-focused) models, "Muse Spark" is a closed model. The new model is available starting today on the Meta AI app and website, and it will later be applied to Meta's three major platforms—Facebook, Instagram, and WhatsApp—as well as its AI-powered smart glasses lineup.
Based on performance benchmarks released by Meta, "Muse Spark" matched the performance of OpenAI's "GPT-5.4," Google's "Gemini 3.1 Pro," and Anthropic's "Claude Opus 4.6" in some areas, and it scored higher than xAI's "Grok" in most categories. The company acknowledged, however, that coding performance is relatively weak.
The "CharXiv Reasoning" metric, which measures chart comprehension, was 86.4%, the highest among the models compared, and the "MMMU-Pro" score, which measures multimodal recognition, was 80.4%, on par with rival models. The coding metrics "SWE-Bench Verified" (77.4%) and "SWE-Bench Pro" (52.4%) were slightly lower than competitors' scores.
"Muse Spark" is the first model in the "Muse" lineup that Meta will roll out going forward. Meta said, "Muse Spark is the first step in MSL's journey, and we are currently developing a larger model."
After the existing "Llama" series was judged to fall short of expectations, CEO Zuckerberg set a goal of achieving "superintelligence," invested massive funds last year, and recruited AI talent—including Scale AI CEO Alexander Wang—to form MSL, a dedicated AI organization.
Alexander Wang, chief AI officer (CAIO) leading MSL, said during an internal Q&A session in Dec. last year that the lab was developing a text-based large language model (LLM) with the codename "Avocado" and an image-and-video-focused model with the codename "Mango." The "Muse Spark" released this time is the model that was called "Avocado" before its public unveiling.