Meta Muse Ember 1.0 Release Notes

Meta has released Muse Ember, a new model designed specifically for high-frequency agentic tasks such as tool calling, coding subtasks, routing, and long-document parsing. By focusing on efficiency, Muse Ember aims to provide a high intelligence-to-cost ratio, positioning itself as a cost-effective alternative to models like Claude Haiku 5.5 and GPT-6 Luna.

Performance and Efficiency

Muse Ember is optimized for the "efficiency frontier," delivering high performance on agentic benchmarks while maintaining low operational costs. According to data from Artificial Analysis, Muse Ember achieves an intelligence score of 36 at a cost of approximately $0.14 per task. This performance is competitive with larger models, often outperforming them in specific agentic harnesses.

Key performance highlights include:

• AutomationBench-AA: Achieved a 37.8% score, significantly outperforming GPT-6 Luna (20.2%) and Claude Haiku 5.5 (13.9%).
• Personal Superintelligence Index: Scored 46.8%, leading over both Luna (37.0%) and Haiku (36.1%).
• Terminal-Bench 4.0: Scored 40.9%, demonstrating strong capabilities in agentic terminal coding.

Technical Specifications and Availability

Muse Ember is available via the Meta Model API as muse-ember-1.0. It features a 1 million-token context window and supports up to 512,000 output tokens. Pricing is set at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens.

The model is natively multimodal, capable of processing text, images, audio, and video without requiring separate specialized models. This allows for direct integration into workflows involving screenshots, scanned documents, and meeting recordings.

Community and Industry Perspectives

The release has prompted significant discussion regarding the current state of model benchmarking and deployment. While the provided metrics show strong performance, some community members have emphasized the importance of workload-specific evaluations over aggregate benchmark scores.

"This reinforces the case for custom, workload specific evals. Broad benchmarks have been useful for shortlisting models but choosing between them, calls for representative tasks, explicit success criteria and measurements using the tools and reasoning settings you actually intend to deploy." — @ProblemExplorer

Other discussions have centered on the lack of open weights, with some users expressing a preference for locally runnable models. Additionally, observers noted that while Muse Ember performs well on specific benchmarks, real-world utility remains the primary metric for adoption among developers who currently rely on established models like Haiku or Luna for software engineering tasks.

Strategic Positioning

Muse Ember is the first model trained using Meta's larger "Muse Wonder Full" architecture. It is intended to round out Meta's current model portfolio, which includes Muse Wonder for high-complexity reasoning and Muse Spark for specialized coding workflows. By offering a lower cost-per-task, Meta aims to make agentic workflows more scalable for enterprise applications, such as contract analysis and candidate screening, as demonstrated by early partners like Box.