Meta releases Muse Ember, a new lightweight model for agent execution

Meta Superintelligence Lab released a caller AI exemplary connected Monday called Muse Ember, the lab’s lightest and astir cost-efficient AI exemplary yet. In its release, Meta said Ember was “designed for analyzable reasoning execution and the enactment agents repetition each day: instrumentality calls, coding subtasks, routing, extraction, and agelong papers parsing.”

Meta peculiarly emphasized Ember’s cost-per-task, utilizing Artificial Analysis’s composite benchmark score to amusement Ember arsenic cheaper and smarter than competitors similar Anthropic’s Haiku and OpenAI’s Luna.

Image Credits:Meta Superintelligence Labs

At 10 cents per cardinal input tokens, Ember is importantly cheaper than frontier competitors similar Haiku and Luna. Not lone that, it’s cost-competitive with open-source competitors similar DeepSeek V4.1 Flash and GLM 5.3 Flash. 

Given the accent connected debased outgo and accuracy astatine scale, Ember is designed to brace with a larger exemplary similar Meta’s Muse Wonder successful an agentic harness. In the lawsuit of a customer-support agent, for instance, you mightiness spot a heavier exemplary establishing rules for however to grip a peculiar issue, past sending a lighter-weight exemplary to use the rules crossed tens of thousands of cases. As Meta puts it successful the announcement, “Muse Wonder plans, Muse Ember handles the volume.”

With volume-heavy usage cases successful mind, Meta besides trained Ember to enactment good crossed antithetic types of media, including text, images, and audio. The effect makes the exemplary amazingly flexible successful what benignant of inputs it tin handle. As Meta puts it successful the announcement, “agents tin work screenshots, charts, and scanned documents, oregon enactment from dependable memos, gathering recordings, and video without a abstracted model.”

Ember is the 4th exemplary released nether Meta’s Muse brand, pursuing the mid-weight Spark (released successful April), Spark’s open-source duplicate Glimmer (released successful August), and the Wonder frontier exemplary (released successful October). The flagship user AI product, the Muse Agent, was released successful September to unexpected success, with caller features similar Spaces and video chat rolling retired earlier this month.