Meta introduces Muse Ember to power cost-efficient AI agents

Meta Platforms Inc. unveiled Muse Ember, a new entry-level artificial intelligence model that promises to be more cost-effective than its competitors. This AI model is the first that Meta has trained using Muse Wonder, the company's leading large language model. While Meta did not specify the exact process, it is likely that Muse Wonder was employed in post-training, which fine-tunes a model's capabilities by further training it with additional data.

Anthropic recently showcased how even smaller language models can perform post-training end-to-end. Meta's Anthropic PBC also recently revealed that future versions of their AI models could significantly speed up machine learning research.

Applications often utilize front-line LLMs, such as Meta's Muse Wonder, as part of AI ensembles. When a user submits a request, these front-line LLMs break it down into smaller tasks and assign them to lower-cost agent models. Muse Ember is specifically engineered for this role. To assess the model, Meta used the Artificial Analysis Index, a tool that combines nine AI benchmarks covering various use cases such as coding and financial analysis.

The evaluation showed that Muse Ember matched Anthropic's entry-level Claude Haiku 5.5 model in performance, but at a much lower cost. It also outperformed OpenAI's GPT-6 Luna algorithm in the same internal tests.

In comparing Muse Ember to its rivals, the model excelled in business task automation, with a 17% lead over the AutomationBench-AA test, which consists of 657 tasks related to popular cloud applications like Gmail and Salesforce. Muse Ember boasts a context window of 1 million tokens, allowing it to process extensive business documents.

Users can also include images, audio, and video in their prompts. The model generates responses up to 512,000 tokens long. Meta offers Muse Ember through its Model API at a rate of 10 cents per million input tokens and 50 cents per million output tokens. If users employ the built-in cache feature, which reuses frequently used prompts to prevent redundant processing, the pricing reduces to 0.4 cents per million input tokens.