Meta is rolling out Muse Spark 1.2, its updated coding model, alongside a beta version of Muse Code, a terminal-based coding agent. The company says it trained the model and the agent together so they'd fit each other's workflows.
Muse Code can plan changes across large codebases, write code, and verify the results. For bigger jobs, it splits the work into subtasks and hands them to separate agents running in isolated workspaces — so your current workspace stays untouched.
Every model call, tool execution, and file edit gets written to a local event log first. That means Muse Code can pick up where it left off after a crash, no need to retype your prompt.
Meta also put Muse Spark 1.2 through a 24-hour optimization run on Nvidia Hopper GPU kernels. It logged more than 1,000 tool calls during the session and was still finding performance gains in the second half.