Muse Ember Operational

Muse Ember is Meta's smallest and most efficient AI model, designed for repetitive agent tasks. It achieves Claude Haiku 5.5's top benchmark scores at about one-third the cost, comprehensively outperforms GPT-6 Luna, and ranks on the Artificial Analysis efficiency frontier. Muse Ember offers a 1-million-token context window and up to 512,000 output tokens, natively supports text, images, audio, and video, and is the first model trained by Muse Wonder Full.

Agent workflow execution: Completes business processes end to end by calling tools in sequence and updating systems across applications, covering repetitive steps such as record retrieval, routing, and information extraction.
Large-scale coding subtasks: Designed for hundreds of small tasks within agent sessions, including reading files, running tests, and making targeted fixes.
Ultra-long context processing: Provides a 1-million-token context window and up to 512,000 output tokens in a single request, allowing it to ingest a complete codebase or document set.
Native multimodal understanding: Processes text, images, audio, and video; reads screenshots, charts, and scans; handles voice memos, meeting recordings, and video; and operates computers directly from screenshots.
Cost-effective reasoning: Reaches Claude Haiku 5.5's top benchmark scores at about one-third of its per-task cost, with each AA task costing about $0.14, low enough to embed in any product.
Enterprise-grade safety: Evaluated by Meta's superintelligence scaling framework before release, covering chemical and biological risks, cybersecurity, loss-of-control risks, and agent behavior safety, all below high-risk thresholds.

Efficiency through scale: Muse Ember is Meta's smallest model to date and takes “intelligence density” to the extreme. It sits on the efficiency frontier across the full Artificial Analysis leaderboard: no model achieves a higher score at a lower per-task cost. Muse Ember deliberately avoids competing directly with flagship models and instead optimizes for the “high-frequency small tasks” that agents perform every day, such as tool calls, coding subtasks, routing, extraction, and long-document parsing. Assigning these tasks to a low-cost small model provides the greatest cost leverage.
Distilled from Muse Wonder Full: Muse Ember is Meta's first specialized model trained through distillation from the flagship Muse Wonder Full. It inherits the flagship model's reasoning and multimodal capabilities, then is streamlined and specialized for efficient execution. Meta's model portfolio is divided accordingly: Muse Wonder handles the hardest problems, Muse Spark handles coding and agent workflows, and Muse Ember handles repetitive tasks at scale and low cost.
Unified long-context and multimodal architecture: The model natively and consistently understands text, images, audio, and video, avoiding the orchestration overhead of calling multiple specialized models for multimodal tasks. Combined with 1 million input tokens and a 512,000-token output window, it can process a complete codebase or document collection in a single request, enabling agents to complete hundreds of coding subtasks continuously within one session without repeated round trips.
Flagship-level safety training: Despite its smaller size, Ember receives the same safety training and testing as core models. Before release, it was evaluated under Meta's superintelligence scaling framework for chemical/biological, cybersecurity, and loss-of-control risks, all below high-risk thresholds. Its behavior when “acting on behalf of users” was also evaluated specifically.

Call it through the Meta Model API: Register on the developer platform at dev.meta.ai and create an API key, then send requests to https://api.meta.ai/v1/responses with the model ID set to muse-ember-1.0. It is compatible with the Responses API format.
Integrate with the OpenAI SDK: Ember is compatible with mainstream SDKs. Point the base URL to https://api.meta.ai/v1 and enter your own key to replace the model seamlessly in existing code.
Program with the Muse Code terminal: Run curl -fsSL https://dev.meta.ai/install.sh | sh to install the official terminal agent, then let Ember automatically read files, run tests, modify code, and complete other tasks from the command line.
Call it through cloud platforms: Select muse-ember-1.0 in Google Cloud Model Garden or Oracle Cloud, then enable and pay for it through each platform's standard process. This is suitable for enterprise users with existing cloud accounts.
Connect through third-party gateways: Purchase usage credits through OpenRouter (meta/muse-ember) or the Vercel / Netlify AI Gateway. A single account can manage multiple models centrally, making this suitable for developers who need to compare and switch between models.

Exceptional cost efficiency: Achieves Claude Haiku 5.5's top benchmark scores at about one-third of its per-task cost, beats all GPT-6 Luna tiers, and remains on the Artificial Analysis efficiency frontier.
Major lead on agent tasks: Scores 37.8 on AutomationBench-AA, nearly 2.7 times Haiku 5.5's score and 1.9 times GPT-6 Luna's score.
Ultra-long context: Supports 1 million input tokens and up to 512,000 output tokens, allowing a complete codebase or document set to be processed in one request and supporting hundreds of coding subtasks in a single session.
Native multimodality: Handles text, images, audio, and video in one model without relying on other models. It can read screenshots, operate computers, and parse scans and meeting recordings, reducing agent orchestration complexity.
Flagship capabilities at smaller scale: As the first model trained by Muse Wonder Full, it inherits the flagship model's reasoning and multimodal capabilities in a minimal form factor.

Official website: https://research.meta.ai/blog/muse-ember

Intelligent enterprise document review: Box uses it to compare contracts line by line against a company's negotiation playbook and provide an overall risk rating, or to screen candidates against detailed resumes and flag borderline-qualified applicants, enabling fast, token-efficient enterprise content processing.
Large-scale coding subtasks: Ember handles hundreds of frequent small tasks in agent sessions, including reading files, running tests, and making targeted fixes, at low cost, while Muse Wonder plans the most difficult architectural designs.
End-to-end business workflow automation: Automatically completes full cross-application business workflows, including retrieving records, calling tools in sequence, and updating systems, covering the repetitive data transfer and process execution common in daily enterprise operations.
Multimodal content processing: Natively processes voice memos, meeting recordings, video, screenshots, and scanned documents, extracting information from unstructured multimedia and performing follow-up actions without connecting to other models.