Muse Ember is out, Meta's small model for agent work

Meta released Muse Ember today. The API name is muse-ember-1.0. It is on Muse Code and the Meta Model API.

This is the small model for work agents do hundreds of times in a session: tool calls, coding subtasks, routing, extraction, and long document parsing. The context window is 1M tokens (a token is a small unit of text). Output goes up to 512K tokens. Price for prompts up to 200K tokens: $0.10 per million input and $0.50 per million output. Cached input is $0.004. Longer prompts cost more.

Meta says Ember is the first model trained using Muse Wonder Full. Here is the setup I am testing. In Muse Code, a Wonder Full lead lists the PDFs and writes a small script that sends each PDF to muse-ember-1.0 through the Model API. Ember workers, set in agents.json, mark each file keep, skip or duplicate, and write Markdown for the keepers. The lead checks the results. I retain each Markdown file in Hindsight as one document with a fixed document_id. Hindsight extracts the facts and stores them as memories.

Primary source: Meta research blog. Docs: Choose a model, Custom agents and Organize photos with subagents. Hindsight: Ingest Data. Coverage: TechCrunch.