Meta Introduces Muse Code Terminal Agent and Muse Spark 1.2 Model

AIDeveloper44 Team

Meta AI Research Releases Muse Code and Muse Spark 1.2

On August 5, 2026, Meta AI Research announced the beta release of Muse Code, a terminal-based coding agent designed for complex software engineering tasks. The agent is powered by Muse Spark 1.2, an updated model specifically optimized for code generation, debugging, and repository-level reasoning. This release follows Muse Spark 1.1 and represents a scaling of training compute and environment diversity for Meta's coding-focused AI ecosystem.

Muse Code is currently available for installation on macOS and Linux systems via a shell command script. The tool is designed to manage large repositories by planning code changes, executing edits, and validating results with minimal user intervention. It operates through a central agent loop supported by a specialized architecture of subagents.

Architecture of Muse Code and Persistent Agents

A primary feature of Muse Code is the use of asynchronous background agents. Unlike traditional agent architectures that may spawn a new process for every task, Muse Code employs persistent background agents that remain active throughout a session. This design is intended to reduce latency by avoiding redundant information gathering. These agents choose when to communicate back to the main agent loop, facilitating progress on multi-step tasks without constant user steering.

The runtime of the agent is built around a local event log. Every model call, tool execution, edit, and user approval is appended to this log. This approach ensures that the system is "replay-exact" and "restart-safe." In the event of a system crash or power failure, the agent can resume its task from the exact point where it stopped, which is critical for long-running software engineering projects that may take hours to complete.

Bundled Commands and Multimodal Capabilities

Muse Code includes several default commands to assist developers in structuring their workflow. The /plan command generates an approval-gated roadmap for a specific task. The /grill command is used to stress-test these plans to identify potential failures before code is written. Finally, the /goal command directs the agent toward the completion of a defined objective. Meta researchers demonstrated the agent's multimodal capabilities by inputting an mp4 video of a home fly-through into the terminal; the agent interpreted the video and generated a functional vacation home marketing and booking page based on the visual data.

Enhancements in Muse Spark 1.2

The Muse Spark 1.2 model serves as the intelligence layer for Muse Code. It is a direct update to Muse Spark 1.1, with expanded training compute focused on coding tasks. To ensure high usability, Meta co-trained the model alongside the Muse Code toolset. This process included rejection-sampled trajectories and recipe optimizations for managing subagents and context compaction.

The model was trained on long-horizon coding tasks, including whole-repository generation and automated research projects. To maintain progress over these extended durations, Muse Spark 1.2 utilizes goal conditioning and context compaction to keep relevant knowledge within its active window. Furthermore, Meta utilized a self-improvement loop where the previous iteration, Muse Spark 1.1, generated challenging coding environments and templates. Muse Spark 1.2 was then trained on solutions graded against these requirements, improving its ability to follow complex, multi-step instructions.

Performance and Case Study: GPU Kernel Optimization

Meta AI Research highlighted the model's performance through a case study involving iterative GPU kernel optimization. In this test, the model was tasked with improving kernel performance over more than 1,000 tool calls, a process that lasted up to 24 hours. The agent wrote, compiled, and profiled code to improve kernels relative to a baseline implementation.

Specifically, the benchmarks targeted KDA and MLA kernels for NVIDIA Hopper GPUs. The agent was prohibited from importing existing third-party kernel libraries; instead, it had to apply knowledge of the Triton programming language to implement optimizations from scratch. For the KDA kernels, Muse Spark 1.2 implemented a chunk-parallel preparation kernel with a sequential inter-chunk scan. For the MLA kernels, it designed a two-kernel Triton pipeline that utilized kernel fusion and tiling, reusing the shared KV latent as both K and V to optimize memory usage and performance.

Availability

Muse Spark 1.2 is currently available through Muse Code and the Meta Model API with global access. Developers can access the beta version of Muse Code to begin integrating agentic workflows into their local development environments on supported platforms. Meta indicated that further updates, including more powerful models and additional harness features, are currently in development.

Enjoyed this?

Get more posts like this delivered to your inbox.

Subscribe