Meta ships Muse Code, a terminal coding agent built on its new Muse Spark 1.2 model

Meta’s research arm has released Muse Code, a beta command-line coding agent, alongside Muse Spark 1.2, the model that powers it. The agent targets large-repository engineering work — planning changes, writing code, and validating results — and leans on persistent background subagents that stay alive across a session instead of being spawned per task, which Meta says cuts redundant context-gathering and the need for constant human steering. A local append-only event log records every model call, tool run, and edit, giving the runtime deterministic replay and the ability to resume mid-task after a crash. The agent ships with built-in skills for generating approval-gated plans, stress-testing them, and driving toward a stated goal.

Muse Spark 1.2 is a coding-focused revision of 1.1, trained with substantially more compute on coding tasks and a wider spread of training environments. Meta co-trained the model with the Muse Code harness itself — feeding in rejection-sampled harness trajectories and tuning for goals, context compaction, and subagent use — so the two are optimized to run together. The team also used the prior 1.1 model to synthesize hard coding environments and then grade candidate solutions, a self-improvement loop meant to sharpen instruction-following.

The headline demonstration is endurance: Meta ran the model on GPU kernel optimization for over 1,000 tool calls and up to 24 hours, iteratively writing, compiling, profiling, and improving Triton kernels for NVIDIA Hopper GPUs. Crucially, the model was barred from importing existing kernel libraries and had to implement the algorithms from scratch. Muse Spark 1.2 is available now inside Muse Code and through Meta’s Model API, with Meta signaling larger, more capable models are on the way — a clear push into the increasingly crowded field of long-horizon autonomous coding agents.