Meta AI releases Muse Code and Muse Spark 1.2

Meta AI released Muse Code (beta) on August 5, 2026, a terminal coding agent built on top of Muse Spark 1.2, its latest coding-focused model. Muse Code can plan, write, and validate code across large repositories while coordinating multiple persistent async background agents — named Photon, Sphere, Embervault, Avo, and Lawn — that remain active throughout a session to reduce latency and redundant data gathering. The runtime uses a local event log that records every model call, tool run, and edit, making the agent crash-resilient and capable of resuming long-running tasks precisely where they stopped. Bundled skills include /plan, /grill, and /goal for structured task execution. Muse Spark 1.2 was co-trained with Muse Code using rejection-sampled trajectories and recipe optimizations, with significantly scaled training compute on coding tasks and expanded training environment diversity. The model was trained on long-horizon tasks including whole-repository generation and large end-to-end projects, and incorporates a self-improvement loop where Muse Spark 1.1 generated challenging coding environments used to train its successor. In benchmark tests, the model demonstrated the ability to iteratively optimize GPU kernels over 1,000+ tool calls across sessions of up to 24 hours. Meta noted larger, more capable models are in development.