Muse Ember is Meta's smallest and most efficient Muse model, released in November 2026 for the many small, repeated steps AI agents take.
In plain words
Most of an agent's tokens go on short, structured steps like calling tools, routing requests and pulling details from documents. Muse Ember is built for that high-volume work and accepts multimodal input, meaning more than just text. It can read up to 1 million tokens in one call,. Meta says that, per dollar spent, it is smarter than both Claude Haiku 5.5 and GPT-6 Luna. On the Meta Model API it costs $0.10 for each million tokens it reads and $0.50 for each million it writes.
A quick example
Imagine a coding agent that hands each small chore, like renaming a function across files or summarising a log, to a cheaper helper model. Muse Ember is meant to be that helper.
Where you'll meet it
- Sub-agents inside coding tools.
- Business workflows that pull records, call tools and update systems.
Related
Sources
- DocsIntroducing Our Smallest, Most Efficient Model: Muse Ember, Meta AI Research · read 9 Nov 2026
- DocsMeta AI for developers, Meta · read 9 Nov 2026
Checked by WhatIsAI against the sources above · 9 Nov 2026. Spotted an error? Report it.