Meta Superintelligence Lab released Muse Ember on Monday, TechCrunch reported, describing it as the lab’s lightest and most cost-efficient model in the Muse line¹.

Meta said the model is designed for repeated agent work such as tool calls, coding subtasks, routing, extraction and long document parsing, according to TechCrunch¹.

TechCrunch said Meta priced Ember at 10 cents per million input tokens and used Artificial Analysis’s composite intelligence index to argue the model is cheaper and stronger on a cost-per-task basis than Anthropic’s Haiku and OpenAI’s Luna¹.

The same report said Ember is also cost-competitive with open-source systems including DeepSeek V4.1 Flash and GLM 5.3 Flash¹.

Meta positioned Ember to run with a larger model such as Muse Wonder in an agent setup, TechCrunch said, for example using a heavier model to set customer-support rules and a lighter one to apply them at scale¹.

In the announcement quoted by TechCrunch, Meta said, “Muse Wonder plans, Muse Ember handles the volume”¹.

TechCrunch said Ember was trained across text, images and audio so agents can work from screenshots, charts, scanned documents, voice memos, meeting recordings and video without a separate model¹.

Ember is the fourth model under the Muse brand, after mid-weight Spark in April, open-source Glimmer in August and the Wonder frontier model in October, TechCrunch reported¹.

The consumer Muse Agent launched in September, the outlet said, with Spaces and video chat features rolling out this month¹.

TechCrunch did not report output-token prices, regional availability or an independent audit of the Artificial Analysis comparison¹.

Sources

1. TechCrunchMeta releases Muse Ember, a new lightweight model for agent execution1:00 PM

Read the original at TechCrunch ↗