Meta Muse Code and Muse Spark 1.2: A New CLI Coding Agent, Explained
Meta launched Muse Code, a terminal coding agent, and Muse Spark 1.2, a code-focused model. Here's how it stacks up on the DeepSWE benchmark.
What is Meta Muse Code?
Muse Code is Meta’s new terminal based coding agent, released in beta and announced by Mark Zuckerberg. It works like Claude Code, Codex CLI, or Grok CLI: you run it from your command line, give it instructions, and it writes and edits code directly in your project. Alongside it, Meta released Muse Spark 1.2, a code-focused model built to power that agent, rather than a general-purpose chat model like Llama.
TL;DR
- Meta released Muse Code, a CLI coding agent similar in concept to Claude Code, Codex CLI, and Grok CLI, meant to run in your terminal and write code on command.
- The agent is paired with Muse Spark 1.2, a model Meta built specifically for coding tasks rather than general chat or reasoning.
- On the DeepSWE benchmark, a test many developers now treat as the clearest signal of real-world coding ability, Muse Spark 1.2 scored 59.3%.
- That score puts it just behind GPT 5.6 Turbo and Opus 5, but notably it wasn’t compared against the current top performers, GPT 5.6 Sol or Fable 5, in Meta’s own materials.
- Meta shared demo outputs including a simple game and a designed website, both of which looked comparable in style to what GPT and Grok models typically produce.
- The launch arrived the same week as Qwen 3.8 Max, a 2.4 trillion parameter open-weight model, giving developers two new coding-capable releases to weigh in a single week.
- For teams choosing a coding agent today, Muse Code enters a market where DeepSWE scores above 70 already exist, meaning Meta’s debut is competitive but not state of the art.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How does Muse Code work?
Muse Code follows the same basic pattern as other terminal coding agents that have become common over the past year. You install it, run it from a command line inside your project directory, and describe what you want built, fixed, or refactored. The agent reads your codebase, makes edits, and can iterate based on feedback, all without leaving the terminal.
This CLI-first approach has become a standard shape for coding agents because it fits directly into how developers already work: inside a repo, inside a shell, without switching to a separate chat window or IDE plugin. Claude Code popularized this pattern, and OpenAI’s Codex CLI and xAI’s Grok CLI followed with similar designs. Meta’s entry with Muse Code signals that the company sees terminal-native agents as a durable interface, not a passing trend.
What is Muse Spark 1.2, and how is it different from Llama?
Muse Spark 1.2 is the model driving Muse Code, and Meta built it specifically for coding rather than as a general-purpose assistant. That distinction matters. General models like Llama are trained to handle a wide range of tasks: writing, reasoning, summarization, conversation. A code-focused model like Muse Spark 1.2 is tuned more narrowly, which can translate into stronger performance on programming tasks even if the model would perform less impressively on unrelated benchmarks.
This is a common strategy right now. Rather than building one model to do everything well, companies are increasingly shipping specialized variants optimized for a single domain, then pairing that model with a purpose-built tool (in this case, a CLI) designed around its strengths.
How does Muse Spark 1.2 score on the DeepSWE benchmark?
DeepSWE has emerged as one of the benchmarks developers pay closest attention to when judging real-world coding ability, more so than broader knowledge or reasoning tests. It’s meant to reflect how a model performs on actual software engineering work rather than isolated coding puzzles.
Muse Spark 1.2 scored 59.3% on DeepSWE. For context:
- Opus 5 and GPT 5.6 Turbo scored slightly higher, placing Muse Spark 1.2 just behind both.
- GPT 5.6 Sol and Fable 5, currently the highest performers on this benchmark, were not included in Meta’s own comparison charts.
- Qwen 3.8 Max, released the same week, scored 56.6%, meaning Muse Spark 1.2 edges out one of its most direct open competitors while still trailing the top tier of closed models.
The omission of GPT 5.6 Sol and Fable 5 from Meta’s benchmark comparisons is worth noting. Companies routinely select comparison points that favor their own release, and leaving out the current benchmark leaders suggests Muse Spark 1.2 wasn’t built to challenge the absolute top of the leaderboard, but rather to compete in the tier just below it.
Is Muse Code worth using instead of Claude Code or Codex CLI?
For developers already relying on Claude Code, Codex CLI, or similar tools built on top of Fable 5 or GPT 5.6 Sol, Muse Code’s underlying model currently sits below those options on the benchmark most relevant to coding work. If raw coding capability is the priority, the higher-scoring models are likely to produce fewer errors and handle more complex engineering tasks with less back and forth.
That said, benchmark position isn’t the only factor that matters. Pricing, rate limits, latency, and how well an agent integrates into a specific workflow all shape whether a tool is genuinely useful day to day. Meta hasn’t published detailed pricing or availability terms beyond the beta launch itself, so a full cost-benefit comparison isn’t yet possible. Developers curious about the tool are best served by testing it directly against their own coding tasks rather than relying on a single benchmark number.
The demo outputs Meta shared, a small game and a styled website, look comparable in visual polish to what GPT and Grok models typically generate for similar prompts. That suggests Muse Spark 1.2 is a capable model for common web and app development tasks, even if it isn’t leading the pack on harder software engineering benchmarks.
Why did Meta release a coding agent now?
The terminal coding agent space has grown crowded fast. Anthropic, OpenAI, and xAI all ship CLI tools tied to their flagship coding models, and open-weight releases like Qwen’s have started closing the gap on benchmarks like DeepSWE. Meta entering this space with both a harness (Muse Code) and a dedicated model (Muse Spark 1.2) suggests the company wants a foothold in developer tooling, not just in chat assistants or open-weight releases like Llama.
It also reflects a broader pattern: coding has become one of the clearest, most measurable use cases for large language models, and benchmarks like DeepSWE give companies a straightforward way to market progress. Releasing a purpose-built coding model alongside a CLI, rather than folding coding ability into a general assistant, mirrors what Anthropic and OpenAI have already done.
Frequently Asked Questions
What is Meta Muse Code?
Muse Code is Meta’s terminal based coding agent, released in beta. It runs from the command line and writes or edits code based on developer instructions, similar to Claude Code or Codex CLI.
What is Muse Spark 1.2?
Muse Spark 1.2 is Meta’s new code-focused model, built to power Muse Code. Unlike Llama, it’s tuned specifically for programming tasks rather than general-purpose conversation or reasoning.
How does Muse Spark 1.2 compare to other coding models?
It scored 59.3% on the DeepSWE benchmark, placing it just behind Opus 5 and GPT 5.6 Turbo, but behind the current top scorers, GPT 5.6 Sol and Fable 5, which Meta didn’t include in its own comparisons.
Can I run Muse Spark 1.2 locally?
Meta hasn’t published local hardware requirements or confirmed consumer GPU compatibility for Muse Spark 1.2. As a code-focused model tied to the Muse Code CLI, it appears designed primarily for cloud or hosted use rather than local deployment.
Is Muse Code available now?
Yes, Meta released Muse Code in beta, announced directly by Mark Zuckerberg, alongside the Muse Spark 1.2 model that powers it.
