Meta has released Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, its newest model. This release marks a step toward more capable models, with Muse Code designed to take on complex software engineering tasks across large repositories. It can plan changes, write code, validate results, and coordinate multiple persistent subagents for each task to solve difficult problems faster and more accurately with less intervention. Muse Code operates using an agent loop and async background agents that remain active throughout sessions, reducing latency by avoiding redundant information gathering and carrying out next steps autonomously. The runtime uses a local event log that records every model call, tool run, approval, and edit, making it replay-exact and restart-safe after crashes, enabling long-running tasks without failure derailment. Bundled skills include /plan for creating approval-gated plans, /grill for stress-testing plans, and /goal for working toward objective completion. Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. It was trained with significantly scaled-up compute on coding tasks and expanded training environment diversity while maintaining strength in general agent capabilities. The model was co-trained with Muse Code using rejection-sampled harness trajectories and recipe optimizations for goals, compaction, and subagents, integrating the Muse Code toolset for harness compatibility. Extensive training on long-horizon coding tasks, including whole-repository generation and auto-research, enables Muse Spark 1.2 to leverage planning, goal conditioning, and context compaction to sustain progress. A self-improvement loop used Muse Spark 1.1 to generate coding environments and instruction-following templates, then graded candidate solutions to build a scalable training dataset, improving instruction-following precision. In a case study, the model optimized GPU kernels over 1,000+ tool calls (up to 24 hours) using Muse Code’s agentic environment, writing, compiling, profiling, and improving kernel performance relative to a baseline FLA Triton implementation for KDA on NVIDIA Hopper GPUs, without importing third-party libraries. Muse Spark 1.2 is available today in Muse Code and via the Meta Model API with expanded global access.
Key facts
- Muse Code is a terminal coding agent powered by Muse Spark 1.2
- Muse Code can coordinate multiple persistent subagents for each task
- Muse Code uses a local event log for replay-exact and restart-safe runtime
- Muse Spark 1.2 was co-trained with Muse Code using rejection sampled harness trajectories
- Muse Spark 1.2 was trained on long-horizon coding tasks including whole-repository generation
- In a case study, Muse Spark 1.2 optimized GPU kernels over 1,000+ tool calls
- Muse Spark 1.2 is available in Muse Code and Meta Model API with expanded global access