Meta introduces Muse Ember, its smallest and most efficient AI model to date

Meta has unveiled Muse Ember, the newest and smallest addition to its family of AI models, positioning it as a lightweight, low-cost workhorse for the growing wave of agentic applications. The model, released as muse-ember-1.0, is available today through the Muse Code platform and the Meta Model API at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens, with a 1-million-token context window and up to 512K output tokens. According to Meta, it costs roughly $0.14 per Artificial Analysis benchmark task, a price point the company says is low enough to embed in “every product, for everyone”.

Early partner testing highlighted speed and efficiency as the model's defining traits. Cloudflare CTO Dane Knecht reported that Muse Ember completed every coding test with the lowest latency measured, a p95 of 2.2 seconds, roughly 4x to 13x faster than leading open-weight models, while using 36–48% fewer output tokens. Oracle's testing found the model ran up to 30% faster than a comparable rival at the tail, while Box praised its accuracy on document-heavy tasks such as scoring contracts against a company's negotiation playbook. Meta says Muse Ember is purpose-built for the repetitive subagent work that agents perform all day, like coding and document analysis, complementing its broader portfolio: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and now Ember for high-volume, repeatable tasks.

Notably, Muse Ember is the first model trained using Muse Wonder Full, fulfilling an earlier commitment to leverage Meta's larger models to develop specialized offerings. The company says the release was vetted under its Meta Superintelligence Scaling Framework, with evaluations across chemical and biological, cybersecurity, and loss-of-control risks finding the model below high-risk thresholds. Despite its small size, it reportedly received the same core safety training as larger models. Meta also says it tested the model's behavior when acting on users' behalf, including how it handles sensitive information and resists hidden instructions embedded in web pages and documents.