An update to my earlier Muse models post. Meta Superintelligence Labs has released Muse Ember, a hosted multimodal model meant for the repetitive steps inside agents: tool calls, coding subtasks, routing, extraction and long-document parsing, MarkTechPost reports.
What’s new
- It’s the fourth Muse-branded model, after Spark, Glimmer and Wonder. Meta’s framing: Wonder plans, Ember handles the volume. Ember is the first model Meta trained with Muse Wonder Full.
- 1 million token context, up to 512K output tokens. Text, images, audio and video, including operating desktop apps from screenshots.
- It ties Claude Haiku 5.5 at 36 on the Artificial Analysis Intelligence Index, at about $0.14 per index task versus Haiku’s $0.41. It leads AutomationBench-AA (37.8 vs 13.9 for Haiku) and is weaker on FrontierCode 2.0-AA (27.2 vs 34.9).
- Hosted only (Meta Model API and Muse Code). Weights are closed, so there’s no local or quantized option.
My take
A cheap, long-context worker model is useful if you already live in Meta’s stack. Closed weights mean you can’t run it yourself or audit it. Price and context are the draw; don’t expect it to replace a top coding model.
God Bless, and May the Source Be with You!
