Meta Introduces Muse Code, a New AI Coding Assistant for Developers
The terminal-based AI tool helps developers organize their software projects more efficiently.
1What is Muse Code?
Meta's new terminal-based coding assistant Muse Code is designed to help developers organize their software projects. Trained on the Muse Spark 1.2 model, the tool offers features similar to competitors like OpenAI Codex or Anthropic Claude Code. It can be used for tasks such as planning codebase changes, writing code, and validating results. Currently available as a closed-source cloud service.
2Agent Architecture Functionality
Muse Code operates as a 'agent orchestrator' running in the command line. When a developer starts a task, background sub-agents create a context file. This file helps recover lost context during tasks. The tool logs all actions before execution, and multiple agents can work in parallel in isolated environments. Zuckerberg shared that in one test, six game features were developed simultaneously without conflicts.
3Capabilities of Muse Spark 1.2
Muse Spark 1.2, the foundation of Muse Code, is the third model released in four months by Meta's Superintelligence Labs. The multimodal model can process text, images, video, audio, and PDFs. Running on Nvidia Hopper GPUs, the system made over 1,000 tool calls in 24 hours. Zuckerberg noted that during a kernel optimization test, the model continued finding significant improvements well beyond the initial exploration phase.
4Competitive Performance
Muse Spark was tested against industry standards like Terminal-Bench 2.1 and DeepSWE 1.1. Results show the model performs close to the best in understanding and executing engineering tasks with minimal deviation. When compared to competitors like Opus 5, GPT5.6 Terra, Grok 4.5 and Gemini 3.6, it holds a competitive position. While scores are generally close, Gemini lagged behind in some tests.
5Technical Specifications and Access
Muse Code can be installed via curl command or accessed through OpenRouter. The API-enabled service is priced at $1.25 per million input tokens and $4.25 per million output tokens, offering a 1 million token context window with discount options. Although currently closed-weight, Zuckerberg hasn't ruled out open-sourcing in the future.
6Meta's AI Strategy
Muse Code is among the first products from Meta's Superintelligence Labs established in June 2025. This unit was created to revitalize the company's AI efforts and develop personalized superintelligence with deep reasoning capabilities. The model's first version launched in April 2025, when Meta researchers acknowledged needing to improve its coding abilities.
7Position in the Market
The coding assistant market has grown rapidly in recent years, with Meta being a relatively late entrant. However, Muse Code's multi-agent architecture and terminal integration differentiate it from competitors. Designed particularly for teams working on large codebases, the tool aims to boost developer productivity. Unlike Meta's open-source approach with Llama models, Muse Spark is currently closed-source.
8Future Expectations
While the beta version of Muse Code currently offers limited capabilities, it's part of Meta's ambitious AI plans. Zuckerberg's statements suggest the model will support more task types in the future. Its potential to go open-source has excited the developer community. With this move, Meta aims to popularize AI usage in software development processes.