Meta AI has released Muse Code in beta, a terminal-based coding agent driven by the new Muse Spark 1.2 model. The software is designed to handle complex engineering tasks across large code repositories by planning changes, writing code, and validating results. A suite of asynchronous background agents remains active throughout the session rather than spawning for every individual task. Every model call, tool run, approval, and edit is recorded in a local, append-only event log. Meta describes this as a replay-exact and restart-safe system. Muse Spark 1.2 was co-trained with the harness itself. The company also published a case study demonstrating the system running over 1,000 tool calls across a period of up to 24 hours.
Deployment
Muse Code is available in beta for macOS and Linux. Installation uses a single command: curl -fsSL https://dev.meta.ai/install.sh | bash. Muse Spark 1.2 is available within Muse Code and the Meta Model API, with expanded global access. The launch announcement does not mention downloadable weights, so treat this as a hosted dependency.
- Company level: The API path fits any size. The Muse Code path fits teams already running agents in sandboxes with review 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, test generation, and GPU kernel optimization.
Asynchronous background agents
Muse Code runs a simple agent loop plus a set of asynchronous background agents. These specialised agents remain active throughout each session. They are not spawned for individual tasks, which Meta says avoids redundant information gathering. They carry out next steps and choose when to report back to the main agent. Meta states this persistence reduces latency and steering on difficult, multi-step tasks.
Runtime design
Muse Code uses a local event log. Every model call, tool run, approval, and edit is appended to it. Meta calls this single source of truth replay-exact and restart-safe. After a crash, the agent resumes precisely where it stopped, letting long-running tasks survive failures.
Bundled skills
Three default skills ship with the agent. /plan turns a task into an approval-gated plan. /grill stress-tests that plan until it holds up. /goal works toward successful completion of the specified objective.
What changed in Muse Spark 1.2
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1. Meta reports gains in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. The research team significantly scaled up training compute on coding tasks and expanded environment diversity. The model keeps its strength in other areas, including general agents.
Three important training details:
- 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 goals, compaction, and subagents. The Muse Code toolset was integrated to maximize harness compatibility.
- Long-horizon training: Training covered whole-repository generation, large end-to-end projects, and auto-research. The model uses planning, goal conditioning, and context compaction to sustain progress.
- Self-improvement: Muse Spark 1.1 generated challenging coding environments and instruction-following templates, then graded candidate solutions against those requirements. That produced a scalable training dataset for 1.2.
Evaluation
Meta’s methodology report is unusually specific. Terminal-Bench 2.1 uses all 89 tasks, pass@1 over five attempts. DeepSWE v1.1 covers 113 tasks across 91 repositories and five languages. Meta Internal Coding Bench holds 440 tasks derived from real internal pull requests. Runs execute in isolated Daytona cloud sandboxes. Comparisons include Grok 4.5, Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash, and Kimi K3, each with its own agent product. Meta notes its harness may not be tuned for third-party models. For reference, Meta’s model page lists Muse Spark 1.1 at 80.0 on Terminal-Bench 2.1.
Case study: kernel optimization
Meta tested iterative GPU kernel optimization over 1,000+ tool calls, running up to 24 hours. The model writes, compiles, profiles, and progressively improves kernels against a provided baseline. Benchmarks covered KDA and MLA kernels on NVIDIA Hopper GPUs. For KDA, the baseline is the FLA Triton implementation, with third-party kernel libraries prohibited. Muse Spark 1.2 paired a chunk-parallel preparation kernel with a sequential inter-chunk scan. For MLA, the reference is PyTorch at batch size 1, 64 heads, sequence length 8192, and latent dimension 512. The model built a two-kernel Triton pipeline that reuses the shared KV latent as both K and V.
What it means
Developers using terminal interfaces gain a workflow that does not stop when the main process crashes. The system remembers exactly where it left off. This is useful for tasks that naturally take hours, such as optimising GPU kernels or refactoring large codebases, because the agent does not need to restart from scratch after an interruption.
