Meta Superintelligence Labs has released Muse Code (beta), a terminal-based coding agent powered by the new Muse Spark 1.2 model. Muse Code features async persistent background agents that reduce latency on multi-step tasks, a local event log for replay-exact and restart-safe execution, and bundled skills like /plan, /grill, and /goal. Muse Spark 1.2 was co-trained with Muse Code, with significantly scaled compute on coding tasks, long-horizon training covering whole-repository generation and auto-research, and a self-improvement loop using Muse Spark 1.1 to generate training data. A case study demonstrates the agent iteratively optimizing GPU kernels (KDA and MLA) for NVIDIA Hopper GPUs over 1,000+ tool calls, achieving substantial improvements over baseline Triton implementations. Muse Spark 1.2 is available in Muse Code and via the Meta Model API.
