Meta Releases Muse Code

Meta has released the beta version of Muse Code, a terminal coding agent based on its latest model, Muse Spark 1.2. Muse Code performs planning changes, writing code, and verifying results in large repositories, coordinating multiple persistent background agents.

Key features include:

• Asynchronous background agents: Specialized agents that persist throughout the session gather information and execute subsequent tasks, reducing latency and user intervention.
• Restartable runtime: Records model calls, tool executions, approvals, and edits to a local event log, allowing tasks to resume exactly after failures.
• Bundled skills: /plan creates approval-based plans, /grill validates plans, and /goal drives the completion of specified objectives.

Muse Spark 1.2 is a coding-specialized model that improves code generation, complex debugging, codebase understanding, and end-to-end development workflows. It was co-trained with Muse Code to enhance performance and compatibility in agentic environments, including goal setting, context compression, and sub-agent utilization.

The model has been trained on long-running tasks such as full repository generation, large-scale end-to-end projects, and automated research, and has been tested on GPU kernel optimization tasks involving up to over 1,000 tool calls and lasting up to 24 hours.