Meta Superintelligence Labs launched Muse Image, an image model that can call its own coding and web-search tools mid-generation, and opened an early preview of Muse Video, a companion video model built on the same pretraining base.
Muse Image is available now inside Meta AI, Instagram Stories, and WhatsApp. Muse Video is preview-only with no public release date attached. Both models sit in the top three of the human-preference Arena leaderboard for their tasks, according to Meta's announcement.
Muse Image can run coding and web-search tools while it generates
Muse Image behaves differently from a standard text-to-image model. It operates agentically. Meta says the model refines its own output, iterates on precise edits, and composes a single image from multiple reference inputs without a human re-prompting at each step.
Two tool integrations do the heavy lifting:
Coding tool. Muse Image generates accurate plots and QR codes, and can build animated GIFs, websites with embedded images, and interactive visual games.
Search tool. The model pulls real-time information from the web to improve factual accuracy on knowledge-intensive prompts, rather than relying only on its training data.
Meta also reports an "approximately log-linear scaling relationship" between output quality and the amount of reasoning and tool calls the model makes at inference time. In practice, letting the model think and call tools longer produces better results. The company attributes the self-refinement to reinforcement learning, where the behavior emerged during training. Muse Image also connects to Meta's Muse Spark for joint planning and accepts inline text and image prompts together.
Distribution runs through Instagram, WhatsApp, and Meta AI
Muse Image is live in the Meta AI app and on meta.ai, in Instagram Stories in the US, and in WhatsApp in a limited set of countries. Facebook access is listed as coming soon.
The Instagram hooks are the most concrete production angle. Users can @-mention public Instagram accounts to pull in reference images, save personalized presets inside the app, and generate marketing assets aimed at small businesses. It is the same social-graph distribution strategy behind Meta's other model launches; we covered Meta's SAM 3D release alongside World Labs and Google's Nano Banana Pro.
Muse Video previews native audio, with Meta naming its own gaps
Muse Video shares Muse Image's pretraining base and is coming to creators and Meta AI, with no date given. Meta describes competitive prompt adherence, visual fidelity, temporal consistency, and native audio generation.
The company is unusually direct about where the model falls short. It names audio-video synchronization and physically accurate fast motion as current weak points, the two areas where AI video most often breaks for professional use.
Arena rankings put Muse behind, not ahead
Meta grounds its quality claims in the Arena leaderboard, which ranks models by human-preference Elo scores. On the rankings the company cites, Muse Image places second in text-to-image, single-image editing, and multi-image editing, while Muse Video places third in text-to-video.
Placing second and third means Meta enters an already crowded top tier rather than leading it, a tier that keeps churning through new leaders as image and video models leapfrog each other release to release.
Content Seal watermarks every image, with video labeling planned
Every Muse Image output carries Content Seal, Meta's invisible watermarking system. Meta says the watermark persists through cropping, compression, resizing, and screenshotting, and that a public detection tool is available at meta.ai/identification. Watermarking for Muse Video is planned but not yet shipped.
What Meta left out
Meta disclosed no image resolution, no video clip length, no pricing, no API availability, and no model size or parameter count. For working professionals, that leaves the practical questions unanswered: how large the outputs are, how long the clips run, and whether Muse will reach production tools beyond Meta's own consumer apps.
The distribution, at least, is concrete. Routing an agentic image model straight into Instagram, WhatsApp, and Facebook drops it into Meta's largest consumer apps, and puts the tool-calling approach directly against the image and video models filmmakers already reach for.
