Editorial illustration for Meta AI Launches Muse Code Beta, Powered by Muse Spark 1.2
Meta Launches Muse Code Beta for Terminal Coding
Meta AI Launches Muse Code Beta, Powered by Muse Spark 1.2
Meta AI put out a beta version of Muse Code on Tuesday, a terminal-based coding agent built to run on Muse Spark 1.2, a model the company trained alongside the harness rather than bolting on afterward. The pitch isn't a chat assistant that suggests snippets. It's a system meant to work across entire repositories: planning multi-file changes, writing the code, then checking its own output before handing it back. Meta frames this release as one step on a longer path toward frontier-scale models, with bigger versions already in the pipeline.
Installation is straightforward. macOS and Linux users can get running with a single curl command pulling from dev.meta.ai, and Muse Spark 1.2 is also reachable through the Meta Model API with wider global availability than prior releases. Meta hasn't published downloadable weights for the model, so anyone adopting this is signing up for a hosted dependency, not a local checkpoint.
The company backed up the launch with a kernel-optimization case study involving more than 1,000 tool calls stretched across a 24-hour session, a stress test meant to show the agent can hold state and keep working on a problem far longer than a typical coding assistant session allows.
Meta AI has released 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 next step toward the frontier, with larger models on the way. Muse Code targets complex software engineering across large repositories: it plans changes, writes code, and validates the results.
Why this matters
Meta is late to the terminal coding agent game, and it knows it. Claude Code, Cursor, and OpenAI's Codex CLI already have developers building habits around them, so Muse Code's beta label and "next step toward the frontier" framing read as an admission that Muse Spark 1.2 isn't there yet. The design choices are the interesting part for us: async background agents that persist for a whole session, and an append-only event log for every model call and tool result, suggest Meta is thinking about auditability and long-running tasks on large repos rather than quick snippet generation.
That's a reasonable bet if you're targeting professional engineering teams instead of solo hobbyists. But "significantly scaled up training compute on coding tasks" is a claim we can't verify from a press release, and Meta hasn't published benchmark comparisons against the tools developers already trust. For founders and engineers deciding what to adopt, the session logging and repo-scale planning are worth testing hands-on.
For researchers, the real question is whether Meta's larger models, teased but not shipped, actually close the gap, or whether this stays a beta experiment.
Related Reading
Browse all Meta news →
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Related Topic OpenAI hires 630 ex-Meta staff as ChatGPT memory may turn data into ads
Business & Startups
Related Topic Meta AI Update Pulls From Your Calendar for Daily Briefings
LLMs & Generative AI
From Archives Mistral's Shieldstral matches larger safety models with rewritten unsafe text.
From Archives AI agents created fake identities in latest hacking test
From Archives Meta AI’s Memory Coach Outperforms Constant Recall for Long Tasks
Research & Benchmarks
From Archives Meta's AI Investments Cut Profit 91% Amid New Data Center Deal
Business & Startups
Common Questions Answered
How does Muse Code differ from traditional chat-based coding assistants?
Muse Code is a terminal-based coding agent designed to work across entire repositories rather than just suggesting code snippets through chat. It can plan multi-file changes, write code, and validate its own output before returning results to the developer, enabling complex software engineering tasks at scale.
What is Muse Spark 1.2 and how was it developed differently?
Muse Spark 1.2 is the model powering Muse Code, and Meta trained it alongside the harness rather than bolting it on afterward. This integrated development approach allows the model to be specifically optimized for the terminal-based coding agent's requirements and workflow.
Why does Meta position Muse Code as a 'next step toward the frontier' rather than a final product?
Meta acknowledges that it is late to the terminal coding agent market, with competitors like Claude Code, Cursor, and OpenAI's Codex CLI already established with developer adoption. The beta label and frontier framing suggest that Muse Spark 1.2 is not yet at the level of frontier-scale models, indicating Meta views this as an intermediate step in their development roadmap.
What architectural features does Muse Code implement to improve code generation reliability?
Muse Code uses async background agents that persist throughout an entire session and maintains an append-only event log for every model call and tool result. These design choices enable better tracking, validation, and reproducibility of the code generation process across complex multi-file projects.
Further Reading
- Meta launches Muse Code AI coding agent for macOS and Linux - 9to5Mac
- Meta debuts first AI coding agent to take on Anthropic and OpenAI - CNBC
- Meta Ships Muse Code Coding Agent With Co-Trained Muse Spark 1.2 Model - Unite.AI
- Meta enters the crowded AI coding battle with Muse Spark 1.1 - TechCrunch
- Meta debuts Muse Spark 1.1 model with preview open to developers - Reuters
