MENU

Meta Muse Ember: Low-cost AI model specs and performance vs Spark

ITmedia NEWS Japan
Overview
Meta has released Muse Ember, a lightweight AI model designed for repetitive agent tasks that achieves performance close to Muse Spark 1.3 at approximately 1/17th of the cost. The model supports multimodal inputs and large context windows, targeting high-volume automation and information extraction.
In Depth

Key Findings

Meta’s new Muse Ember model significantly reduces operational costs for repetitive AI tasks. According to the Intelligence Index v5 by Artificial Analysis, Muse Ember scored 36, nearly matching the 37 scored by Muse Spark 1.3. However, the cost per task for Muse Ember is 0.14 dollars, compared to 2.39 dollars for Spark 1.3, representing a reduction to approximately 1/17th of the original cost.

Technical Details

Muse Ember is optimized for tool calling, programming assistance, information extraction, and long-document analysis. Its technical specifications include:

  • Input Support: Text, images, audio, video, and PDFs.
  • Context Window: Up to 1 million tokens.
  • Maximum Output: 512,000 tokens.
  • Pricing: 0.10 dollars per 1 million input tokens and 0.50 dollars per 1 million output tokens via the Meta Model API. Cached inputs are priced at 0.004 dollars per 1 million tokens.

In benchmarks, Muse Ember recorded 37.8% on AutomationBench-AA, surpassing GPT-6 Luna (20.2%) and Claude Haiku 5.5 (13.9%). It also scored 40.9% on Terminal-Bench 4.0, though it performed lower than Claude Haiku 5.5 on the FrontierCode 2.0-AA benchmark.

Background and Context

Muse Ember is the first model trained using Muse Wonder Full. It fits into a tiered ecosystem where Muse Wonder handles complex problem solving and long-term tasks, Muse Spark manages coding and agent processing, and Muse Ember executes high-volume repetitive tasks. Meta envisions a workflow where Muse Wonder creates the overall plan while Muse Ember handles individual program fixes and data collection.

Significance and Outlook

The release follows a trend of lightweight model development from competitors, including OpenAI’s GPT-5.4 nano, Google’s Gemini 3.5 Flash-Lite, and Anthropic’s Claude Haiku 5.5. Muse Ember is available as “muse-ember-1.0” via the Meta Model API and Muse Code, and is also provided through Google Cloud and Oracle Cloud.

Source: https://www.itmedia.co.jp/news/article/2611/10/2000003044/

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.

  • Your email and selected fields are used only to deliver the newsletter.
  • We never share your information with third parties.
  • You can unsubscribe anytime via the link in each email.

See our Privacy Policy for details.

Takes about a minute · Unsubscribe anytime

Published by Troy-Technical, an independent site run by one engineer with a career in materials development.
About the author / Contact info@troy-technical.jp
Let's share this post !

Author of this article

TOC