Meta Superintelligence Lab has unveiled a lightweight AI model called "Muse Ember", TechCrunch reported on Nov. 9 local time.

Ember is the lightest and most cost-efficient among Meta's models, the report said.

Ember runs complex reasoning and handles repeated agent tasks such as tool calling, coding subtasks, routing, information extraction and long-document analysis.

Meta in particular cited an Artificial Analysis aggregate benchmark score to stress that Ember is cheaper than Anthropic Haiku and OpenAI Luna while delivering better performance.

Ember costs 10 cents per 1 million input tokens. It is much cheaper than Haiku and Luna, and is also price-competitive compared with open-source models such as DeepSeek V4.1 Flash and GLM 5.3 Flash, TechCrunch said.

Ember can share workloads with a large model such as Meta's "Muse Wonder" inside an agent harness. Using a customer support agent as an example, a large model sets the rules for handling problems and a lightweight model applies those rules to cases. Meta said, "Wonder plans and Ember handles the volume."

Meta said it trained Ember to handle multiple data formats such as text, images and audio well. It explained that an agent can read screenshots, charts and scanned documents without a separate model, and can also work with voice memos, meeting recordings and videos.

Ember is the fourth model released under the Muse brand. Meta sequentially launched the mid-sized model "Spark" in April, the open-source version of Spark called "Glimmer" in August, and the frontier model "Wonder" in October.