— Muse Code takes on complex software engineering tasks across large repositories: planning changes, writing code, and validating the results, Meta says.
It can coordinate subagents that are persistent and run in the background through each session, instead of respawning for individual tasks, reducing latency and feedback loops.
The agents also use advanced logging of every call, tool run, approval and edit, meaning they should run directly after restarts and crashes, picking up from the latest log entry like nothing happened.
The underlying model, Muse Spark 1.2, scores well on benchmarks for its price point. It clocks in just behind Opus 5’s 86.7% on Terminal-Bench with 82.9%, beating GPT-5.6 Terra and Gemini 3.6 Flash. On DeepSWE it scores 59.3%, ticking in behind GPT Terra and Opus 5 Max. On Artificial Analysis, a more general benchmark, it scores a little behind GPT-5.5.
— Muse Spark 1.2 was extensively trained on long-horizon coding tasks, including whole-repository generation, large end-to-end projects, and auto-research, Meta says on the model.
Muse Code is available with a «one-line install» from the terminal on macOS and Linux, and is available on dev.meta.ai on the web.
Read more: Meta’s launch page, X post by Zuckerberg, X post from Artificial Analysis. Writeups on CNBC, TechCrunch, and Reuters. Discussion on Hacker News and r/Singularity.