Meta has released Muse Spark 1.2 and a dedicated Muse Code agent, presenting the update as a coding-focused improvement over Muse Spark 1.1. The model is trained more heavily on programming tasks and long-running work such as repository generation and independent research.
The Decoder reports that Meta claims gains in code generation, debugging, and reasoning over large codebases. Muse Code includes familiar terminal-agent features such as planning, but also adds a “grill” mode to stress-test plans and persistent sub-agents that can report back during a session. A fast-resume feature logs calls, approvals, and changes so the agent can recover after a crash.
The pricing angle is central. Meta is offering a very cheap output-token tier, but users pay for it with their data. The published benchmarks also carry caveats: competing systems may not have been tuned for the same setup, and some comparisons do not map cleanly to official leaderboards.
The release shows how coding-agent competition is widening from raw model quality to price, workflow features, and data terms.