Meta Releases Muse Code and Muse Spark 1.2: Coding Agent and Model Update with Contributor Version at 95% Off

Meta has released the coding agent Muse Code and the latest large model Muse Spark 1.2. Muse Code is in open beta, installable with one command. Muse Spark 1.2 is the result of training 1.1 with the harness, offering limited improvements but providing a contributor version.

The contributor version requires users to agree to upload data, granting a 0.5% discount, saving 95%, cheaper than DeepSeek. This is interpreted as a strategy of collecting data at a loss, with intentions clear to all.

Muse Code accepts complete engineering tasks, understanding requirements in large codebases, planning changes, making cross-file modifications, and verifying results. In the official demo, a user inputs a house aerial video, and Muse Code generates a vacation home marketing page and verifies it in the browser.

Its structure is a simple main loop plus asynchronous background agents, which remain active throughout the session, advancing on their own and deciding when to report back. The background observer agent works around four dimensions: memory recall, skill recall, goal tracking, and verification. The first three are enabled by default, while verification is off by default, but it incurs additional token overhead.

When tasks can be split, Muse Code dispatches to multiple sub-agents, with isolation enabled via muse -subagent-worktree-isolation. Each sub-agent creates an independent Git worktree from the main commit point, commits independently, and the main agent reviews and merges. A local event log records every model call, tool run, approval, and edit, serving as the single source of truth, supporting precise replay and restart safety.

Muse Spark 1.2 is deeply co-trained with Muse Code, with training material including rejection-sampled harness trajectories, as well as recipe optimizations for goal setting, context compression, and sub-agents. The toolset is directly integrated into training.

Long-horizon capabilities come from whole-repository generation, large end-to-end projects, and automated research tasks. The model uses planning and sequencing, goal conditioning to maintain direction, and context compression to preserve knowledge. A self-improvement loop uses the previous generation 1.1 to generate coding environments and instruction templates, which the model scores, forming scalable training data.

Meta showcased a high-intensity case: optimizing KDA and MLA kernels on NVIDIA Hopper GPUs, executing over 1000 tool calls, running for up to 24 hours. Under the constraint of not importing third-party kernel libraries, the model wrote, compiled, profiled, and improved its own code, achieving significant performance gains over the baseline.