A new terminal agent from Meta takes on whole software engineering tasks with reviewers running in the background. The company calls Muse Code its first coding agent and pairs it with a Muse Spark 1.2 model co-trained to work alongside it.
The agent installs with a single command, then moves from planning changes through writing code and validating the results. Meta says parallel workers handle implementation while reviewers check the output, and that every action is transparent and traceable. Training the model and the agent together, the company argues, sharpens tool handling, cuts retries, and lifts output quality beyond what a generic wrapper around an external model manages.
Muse Spark 1.2 comes with a 1 million-token context window, roomy enough for marathon jobs to stay inside a single session. Meta publishes vendor-reported benchmark charts covering Terminal-Bench, DeepSWE, its internal coding bench, and GDPval without a methodology write-up.
Access runs through the Meta Model API, now in public preview. The standard tier charges $1.25 per million input tokens and $4.25 per million output tokens, with prompts kept out of training data. A contributor tier drops prices to $0.10 and $0.20 per million tokens in exchange for letting Meta use the data to improve its models. Spark 1.2 is also listed on OpenRouter.
The release puts Meta in direct competition with established terminal agents from Anthropic and OpenAI.
