Meta has launched Muse Spark, described as its most capable AI yet and the first model from Meta’s Superintelligence Labs, formed nine months ago under Chief AI Officer Alexandr Wang after Meta’s $14 billion acquisition of Scale AI. Muse Spark is live at meta.ai and on the Meta AI app, with rollout to Facebook, Instagram, and WhatsApp planned in coming weeks; the model is natively multimodal, processing images, text, and voice, and Meta positions it as a response to Google’s Gemini Deep Think and OpenAI’s GPT Pro thinking modes.
Muse Spark is natively multimodal, processing images, text, and voice from the ground up, and it incorporates visual chain-of-thought to make its internal visual reasoning explicit. The model includes tool-use support and features a Contemplating mode in which multiple AI agents work in parallel to pursue extended reasoning and coordinated problem-solving.
Meta described Muse Spark as the first product of its Superintelligence Labs and the first step on Meta’s scaling ladder, and said the model’s capabilities are supported by investments across the stack. Meta stated it is making strategic investments across research, model training, and infrastructure, explicitly naming the Hyperion data center. Meta said these investments support further scaling of Muse Spark’s training and deployment.
Muse Spark has shown strong performance on several benchmarks, particularly excelling on the CharXiv Reasoning test with a score of 86.4, the highest among tested models. It also performed well on HealthBench Hard, scoring 42.8, ahead of GPT 5.4’s 40.1 and significantly above Gemini 3.1 Pro’s 20.6. In DeepSearchQA, Muse Spark scored 74.8, surpassing Gemini’s 69.7 and slightly outperforming GPT 5.4’s 73.6. However, in ARC AGI 2, Muse Spark’s score of 42.5 lagged behind Gemini’s 76.5. Competitive interactions in LiveCodeBench Pro and MMMU Pro also saw Muse Spark trailing Gemini, which leads on these categories with scores of 82.9 and 83.9 respectively against Muse Spark’s 80.0 and 80.4. Meta acknowledges gaps in long-horizon agentic systems and coding workflows.
A system prompt for Muse Spark circulated on social media via a tweet by Pliny the Liberator. Meta described Muse Spark as the first step on its AI scaling ladder and the first product of a ground-up overhaul of its AI efforts. Muse Spark is native multimodal and designed with health-reasoning capabilities, but it does not lead every benchmark or leaderboard. The release and related events underscore Muse Spark’s role within Meta’s broader AI development program. The model represents a significant milestone in Meta’s ongoing investments in AI infrastructure and scaling.
