Meta Unveils Muse Ember Small Language Model to Power AI Agents
Meta has launched Muse Ember, an ultra-fast small language model designed to power autonomous AI agents with high efficiency, massive context windows, and competitive pricing.
Meta has announced the launch of Muse Ember, its smallest, fastest, and most efficient small language model to date, designed specifically to power autonomous AI agents handling hundreds of repetitive daily operations.
Alexander Wang, head of AI at Meta, explained on X that Ember was engineered from the ground up to serve as a tireless digital workforce. While Meta's flagship model Muse Wonder handles high-level project planning, Ember executes the granular sub-tasks: reading contracts, routing support tickets, extracting form data, and running rapid code checks.
Despite its compact footprint, Ember features a context window of one million tokens and supports output generation up to 512,000 tokens per request. As a fully multi-modal model, it processes images, video, and audio without a separate transcription model, and can operate computers via screenshots. On the AutomationBench-AA benchmark, Ember scored 37.8 points, significantly outperforming Claude Haiku 5.5 at 13.9 and OpenAI's GPT-6 Luna.
Industry Feedback and Pricing Strategy
Early enterprise partners have praised the system's performance. Cloudflare reported that Ember runs up to 30% faster than competing models in edge requests, while Box highlighted its precision in contract auditing. Meta is aggressively targeting developers with pricing set at ten cents per million input tokens and fifty cents per million output tokens, bringing the average task cost to just fourteen cents.
However, market competition remains fierce. While Ember's base rate matches OpenAI's GPT-6 Luna, OpenAI offers an aggressive cache discount of roughly 90% on recurring tokens. This means Luna may prove more economical for developers repeatedly running identical system prompts, creating a tight race in the enterprise agent market.
