Meta AI Releases Muse Code (Beta): A Terminal Coding Agent Powered by the New Muse Spark 1.2 Model

Meta AI has launched Muse Code (in beta), a terminal coding agent in beta, powered by its new Muse Spark 1.2 model. Meta positions the pair as its subsequent step towards the frontier, with bigger fashions on the method. Muse Code targets complicated software program engineering throughout giant repositories: it plans adjustments, writes code, and validates the outcomes. A set of async background brokers stays alive for the entire session as an alternative of spawning per job. A native append-only occasion log data each mannequin name, device run, approval, and edit, which Meta calls replay-exact and restart-safe. Muse Spark 1.2 was co-trained with the harness itself. Meta additionally revealed a kernel-optimization case examine operating 1,000+ device calls over so long as 24 hours.

Is it deployable

Yes. Muse Code ships in beta for macOS and Linux through curl -fsSL https://dev.meta.ai/set up.sh | bash. Muse Spark 1.2 is offered in Muse Code and the Meta Model API, with expanded world entry. The launch submit doesn’t point out downloadable weights, so deal with this as a hosted dependency.

• Company stage: The API path matches any measurement. The Muse Code path matches groups already operating brokers in sandboxes with evaluate gates.
• Industries: Software and SaaS, developer tooling, fintech engineering, GPU and inference infrastructure, semiconductors and HPC.
• Applications: Repository-scale refactors and migrations, long-running bug triage, take a look at technology, and GPU kernel optimization.

Async background brokers

Muse Code runs a easy agent loop plus a set of async background brokers. These specialised brokers stay energetic all through every session. They will not be spawned for particular person duties, which Meta says avoids redundant data gathering. They perform subsequent steps and select when to report again to the most important agent. Meta states this persistence reduces latency and steering on troublesome, multi-step duties.

Runtime design

Muse Code makes use of an area occasion log. Every mannequin name, device run, approval, and edit is appended to it. Meta calls this single supply of reality replay-exact and restart-safe. After a crash, the agent resumes exactly the place it stopped, letting long-running duties survive failures.

Bundled abilities

Three default abilities ship with the agent. /plan turns a job into an approval-gated plan. /grill stress-tests that plan till it holds up. /aim works towards profitable completion of the specified goal.

What modified in Muse Spark 1.2

Muse Spark 1.2 is a coding-focused replace to Muse Spark 1.1. Meta reviews positive aspects in code technology, complicated debugging, codebase understanding, and end-to-end developer workflows. The analysis crew considerably scaled up coaching compute on coding duties and expanded surroundings range. The mannequin retains its power in different areas, together with basic brokers.

Three essential coaching particulars:

• Co-training with the harness: Muse Spark 1.2 was co-trained with Muse Code. Training included rejection-sampled harness trajectories and recipe optimizations for objectives, compaction, and subagents. The Muse Code toolset was built-in to maximise harness compatibility.
• Long-horizon coaching: Training lined whole-repository technology, giant end-to-end tasks, and auto-research. The mannequin makes use of planning, aim conditioning, and context compaction to maintain progress.
• Self-improvement: Muse Spark 1.1 generated difficult coding environments and instruction-following templates, then graded candidate options in opposition to these necessities. That produced a scalable coaching dataset for 1.2.

Evaluation

Meta’s methodology report is unusually particular. Terminal-Bench 2.1 makes use of all 89 duties, go@1 over 5 makes an attempt. DeepSWE v1.1 covers 113 duties throughout 91 repositories and 5 languages. Meta Internal Coding Bench holds 440 duties derived from actual inside pull requests. Runs execute in remoted Daytona cloud sandboxes. Comparisons embody Grok 4.5, Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash, and Kimi K3, every with its personal agent product. Meta notes its harness will not be tuned for third-party fashions. For reference, Meta’s model page lists Muse Spark 1.1 at 80.0 on Terminal-Bench 2.1.