Our Smallest, Most Efficient Model: Muse Ember

We are pleased to announce the debut of Muse Ember, our most compact and efficient model tailored for intricate reasoning tasks and frequently repeated agent activities. These include tool calls, coding subtasks, routing, data extraction, and parsing lengthy documents. Muse Ember surpasses Claude Haiku 5.5 and GPT-6 Luna in terms of intelligence per dollar, delivering the same performance at approximately one-third the cost.

Furthermore, it stands at the efficiency pinnacle of the Artificial Analysis leaderboard, with no other model outperforming it for a lower price per task.

For comprehensive evaluation insights, refer to our detailed report. Muse Ember is now accessible via the Meta Model API at a rate of $0.10 per million input tokens and $0.50 per million output tokens. During preliminary testing, our partners utilized Muse Ember for typical agent tasks such as contract review, coding, tool calls, and system updates across various applications.

Its 1M-token context window and 512K output token capacity allow it to handle entire codebases or document sets in a single call. Designed for coding subtasks an agent performs hundreds of times per session, Muse Ember excels in reading files, running tests, making targeted fixes, and working with visual data from screenshots, charts, and scanned documents, or from voice memos, meeting recordings, and videos.

Before its release, Muse Ember underwent rigorous evaluation under the Meta Superintelligence Scaling Framework, assessing chemical, biological, cybersecurity, and loss-of-control risks. Across all categories, it remained below the framework's high-risk thresholds. Despite its modest size, Muse Ember received the same comprehensive safety training and testing as our other models.

We also evaluated its behavior when acting on users' behalf, ensuring it respects user intent, handles sensitive data appropriately, resists manipulation through hidden instructions, and seeks human input before high-impact actions.

Muse Ember is currently available on the Muse Code platform and Meta Model API under the identifier muse-ember-1.0. Priced at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens, it offers a 1M-token context window and 512K output token capacity. It supports the Responses API.

Following the introduction of Muse Wonder and Muse Wonder Full, we aimed to deploy these models for specialized applications. Muse Ember, trained using Muse Wonder Full, marks the first such deployment, expanding our model portfolio to include options for various tasks. Muse Wonder excels in tackling the most challenging problems, Muse Spark is optimized for coding and agentic workflows, and Muse Ember is ideal for repetitive coding and agentic tasks on a large scale at an affordable price of $0.14 per Artificial Analysis task.

This affordable pricing ensures that models of this caliber can be integrated into any product, making superintelligence accessible to everyone. Our ambition extends beyond capability; we strive to make these advanced models attainable at a price that scales, paving the way for a future where everyone can harness the power of personal superintelligence. Further efficient models are in development, and we eagerly anticipate sharing future advancements.