Meta Launches Muse Code Beta With Muse Spark 1.2 for AI Coding

Meta has launched Muse Code, a terminal-based coding agent powered by its new Muse Spark 1.2 model, adding tools for software engineering workflows and long-running development tasks to its AI coding portfolio. 

The company said Muse Code, currently available in beta for macOS and Linux, can plan code changes, write code, validate results and coordinate multiple persistent subagents to work across large code repositories. 

Muse Spark 1.2, the underlying model, is available through Muse Code and the Meta Model API.

“We’re excited to release Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our newest model,” Meta said in a blog post. “This marks our next step towards the frontier, with larger and much more capable models on the way.”

Muse Code runs on a local event log that records every model call, tool execution, approval and code edit, allowing sessions to resume after interruptions. Rather than creating new agents for each task, it relies on persistent background agents that remain active throughout a session to gather information and execute steps with less user intervention.

The coding agent also includes built-in commands such as /plan, which converts a request into an approval-based execution plan, /grill, which tests the plan before execution, and /goal, which works towards completing a specified objective.

Alongside the agent, Meta introduced Muse Spark 1.2, a coding-focused update to Muse Spark 1.1. The company said the model improves code generation, debugging, codebase understanding and end-to-end software development workflows after increasing training compute for coding tasks and expanding the diversity of training environments.

Meta said Muse Spark 1.2 was co-trained with Muse Code using rejection-sampled harness trajectories, optimisations for planning, context compaction and subagent coordination, along with integration of Muse Code’s toolset to improve compatibility between the model and the coding agent.

The company said the model was trained on long-horizon coding tasks, including whole-repository code generation, large software projects and automated research. It uses planning, goal conditioning and context compaction to maintain progress across extended development sessions.

Meta also said it used Muse Spark 1.1 to generate coding environments and instruction-following templates that were then used to train Muse Spark 1.2. The earlier model evaluated candidate solutions to create a larger training dataset, a process the company described as a self-improvement loop.

As part of its evaluation, Meta tested Muse Spark 1.2 on GPU kernel optimisation tasks involving more than 1,000 tool calls over as long as 24 hours. The company said the model wrote, compiled, profiled and iteratively improved NVIDIA Hopper GPU kernels, delivering performance gains over a baseline implementation without relying on third-party kernel libraries.