AI at Meta has announced Muse Ember, what it describes as its smallest and most efficient model, designed for agent workflows including tool calling, code fixes and work across large document sets.
In a post on X, the company said Muse Ember can carry workflows across apps, call tools in sequence, run tests and make targeted code fixes. The post also said it has a 1-million-token context window and native understanding of text, images, audio and video.
AI at Meta said in a post on X that on the @ArtificialAnlys Intelligence Index, Muse Ember performs comparably to Claude Haiku 5.5 at roughly one-third of the cost per task, and scores higher than GPT-6 Luna across all of Luna's reasoning settings.
Analysis
Why This Matters
- Muse Ember targets agent-based AI workflows, a rapidly growing area where models are expected to perform multiple tasks autonomously.
- Its large context window could enable processing of entire document sets in a single session, potentially reducing the need for chunking or retrieval.
- The claimed cost advantage over rivals may pressure other providers to adjust pricing for agent workloads.
Background
Meta has been expanding its AI model portfolio, previously releasing the Llama series of large language models. Muse Ember appears aimed at a different niche: small, efficient models for agent tasks rather than general-purpose chat or generation.
Key Perspectives
AI at Meta: said in a post that Muse Ember is built for complex reasoning execution and everyday agent work, and that it delivers more intelligence for the cost.
What to Watch
- Whether independent benchmarks confirm the claimed performance and cost comparisons against Claude Haiku 5.5 and GPT-6 Luna.
- How developers adopt Muse Ember for production agent workflows given the model's small size and large context window.
- Whether Meta releases an open-weight version, as it has done with Llama models, or keeps Muse Ember proprietary.
Checked
12 claims in this story checked against 2 sources: 11 supported.
Show each claim
- SupportedMeta introduces Muse Ember, a small AI model for agent workflows.AI at Meta on X: “Today we're introducing Muse Ember, our smallest and most efficient model, built for complex reasoning execution and everyday agent work.”
- Partly supportedThe company says Muse Ember outperforms rivals at lower cost.AI at Meta on X: “it achieves performance comparable to Claude Haiku 5.5 at roughly one-third of the cost per task, while scoring higher than GPT-6 Luna across all of Luna's reasoning settings.”
- SupportedAI at Meta describes Muse Ember as its smallest and most efficient model.Meta AI Research: “Muse Ember is our smallest and most efficient model”
- SupportedMuse Ember is designed for agent workflows, including tool use, code fixes and work across large document sets.AI at Meta on X: “Muse Ember can carry workflows across apps, call tools in sequence, run tests and make targeted code fixes. With a 1M-token context window and native understanding of text, images, audio and video, it can also work across large document sets and varied inputs.”
- SupportedMuse Ember can carry workflows across apps.AI at Meta on X: “Muse Ember can carry workflows across apps”
- SupportedMuse Ember can call tools in sequence.AI at Meta on X: “call tools in sequence”
- SupportedMuse Ember can run tests.AI at Meta on X: “run tests”
- SupportedMuse Ember can make targeted code fixes.AI at Meta on X: “make targeted code fixes”
- SupportedMuse Ember has a 1-million-token context window.AI at Meta on X: “With a 1M-token context window”
- SupportedMuse Ember natively understands text, images, audio and video.AI at Meta on X: “native understanding of text, images, audio and video”
- SupportedOn the @ArtificialAnlys Intelligence Index, Muse Ember performs comparably to Claude Haiku 5.5 at roughly one-third the cost per task.AI at Meta on X: “On the @ArtificialAnlys Intelligence Index, it achieves performance comparable to Claude Haiku 5.5 at roughly one-third of the cost per task”
- SupportedOn the index, Muse Ember scores higher than GPT-6 Luna across all of Luna's reasoning settings.AI at Meta on X: “while scoring higher than GPT-6 Luna across all of Luna's reasoning settings.”
