The price-performance arms race in artificial intelligence has welcomed another heavyweight contender. On November 9, Meta officially released Muse Ember, its smallest and most efficient lightweight AI model to date.
Built for High-Frequency Tasks: Meeting Intensive Agent Workloads with Extreme Efficiency
As a core product focused on high efficiency and lightweight design, Meta has precisely targeted Muse Ember's use cases at the high-frequency tasks that AI agents need to handle repeatedly in daily operations. These scenarios specifically include tool calling, executing coding subtasks, request routing, information extraction, and long-document parsing.
In terms of supported input and output formats, the model demonstrates strong multimodal processing capabilities. It can seamlessly accept multiple input types, including text, images, audio, video, and PDFs, and consistently outputs results in text form.
Massive Context Window: Easily Handle Massive Codebases and Long Documents in a Single Request
In terms of architecture design and processing capacity, Muse Ember has an ultra-large context window of about 1 million tokens and can output up to about 512,000 tokens.
This excellent long-context processing capability means it can easily accommodate and process large-scale code repositories or massive document collections in a single request, providing solid foundational support for developers working on complex engineering projects. The model is now officially available through Muse Code and the Meta Model API, under the unified model ID “muse-ember-1.0”.
Authoritative Evaluation Revealed: Performance on Par with Claude Haiku5.5, with Per-Task Cost Sharply Reduced
According to test data published by well-known third-party evaluation firm Artificial Analysis, Muse Ember scored 36 points on the comprehensive intelligence index, exactly matching Anthropic's Claude Haiku5.5 at its highest reasoning tier.
However, when it comes to the closely watched usage cost, Muse Ember demonstrates extremely strong competitiveness. Its official pricing is: $0.10 per million input tokens, $0.004 for cached input, and $0.50 per million output tokens. Artificial Analysis's measured data shows that at its highest reasoning tier, the cost per task is about $0.14, while Claude Haiku5.5 costs as much as about $0.41 to complete similar tasks—bringing the cost down to roughly one-third of the latter.
However, evaluation results across different dimensions and projects also show certain areas of emphasis. According to tests Meta itself has published, Muse Ember outperforms Claude Haiku5.5 and OpenAI's GPT-6Luna on some agent, tool-calling, and multimodal tasks; but in the specialized FrontierCode2.0 test, which examines whether code changes can be smoothly adopted by project maintainers, its score is slightly lower than Claude Haiku5.5.
