Introducing Muse Code (beta), a terminal coding agent powered by the new Muse Spark 1.2 model. This release represents a significant advancement, with larger and more capable models anticipated. Muse Code tackles complex software engineering tasks, including planning, coding, and validation, across large repositories.

It utilizes coordinated persistent subagents for enhanced speed and accuracy, operating with a local event log for replay-exact and restart-safe execution, enabling long-running tasks. Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, featuring improvements in code generation, debugging, and developer workflows, achieved through scaled training compute and diverse environments. It was co-trained with Muse Code for optimal performance.

The model excels in long-horizon coding tasks, leveraging planning, goal conditioning, and context compaction. A self-improvement loop, using Muse Spark 1.1 to generate challenging coding environments, enhanced Muse Spark 1.2's instruction-following precision. In a case study, the model iteratively optimized GPU kernels over 1,000+ tool calls. Muse Spark 1.2 is available today in Muse Code and the Meta Model API.