Summary
- Meta has introduced Muse Code (beta), a terminal coding tool powered by the upgraded Muse Spark 1.2 model, available through the Meta Model API and a curl install script.
- This agent operates with persistent background subagents and maintains a precise event log, allowing it to resume from the exact point of a crash.
- According to Meta's assessments, Muse Spark 1.2 lags behind Anthropic's Opus 5 across all coding benchmarks while outperforming OpenAI's Codex and Google's Antigravity in most cases.
Meta has joined the ranks of tech leaders releasing coding agents, vying for a competitive edge against major AI players like Anthropic and OpenAI.
"We are thrilled to unveil Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our latest model," the company stated in an official announcement. "This represents a significant step towards our future goals, with larger and more capable models forthcoming."
Muse Code is designed as an agentic coding solution for software engineering within extensive repositories. As per Meta, it "handles complex software engineering tasks across large repositories: planning changes, coding, and validating outcomes. It can manage multiple persistent subagents for each task, resolving challenging problems more quickly, accurately, and with less human intervention."
A notable feature is its runtime capability. Muse Code records every model call, tool execution, approval, and modification in a local event log that serves as a definitive record. "This single source of truth ensures the runtime is replay-exact and restart-safe: after a crash, the agent can continue exactly where it left off," Meta explained. For lengthy tasks, this feature is more crucial than mere speed—an aspect competitors have not highlighted.
Muse Code also comes equipped with default functionalities. The "/plan" command transforms a task into a plan requiring approval, while "/grill" rigorously tests that plan until it is robust, and "/goal" guides the task toward successful completion, similar to Hermes. Meta indicated that Muse Spark 1.2 was co-trained with Muse Code to ensure harmony between the core LLM and the agent.
Performance Metrics and Considerations
Muse Spark 1.2 is a coding-centric upgrade from Muse Spark 1.1. Meta reported that it "significantly enhanced training compute for coding tasks and diversified the training environment, leading to advancements in code generation, complex debugging, and comprehensive developer workflows." The performance data is telling.
In Terminal-Bench 2.1, Muse Spark 1.2 with Muse Code achieved a score of 82.9%, trailing behind Claude Code on Opus 5, which scored 86.7%, but surpassed GPT-5.6 Terra on Codex (81.8%) and Grok Build (81.6%).
In terms of agentic coding capabilities as measured by DeepSWE 1.1, Muse scored 59.3%, compared to Opus 5's 65.0% and Codex's 64.8%. On Meta's internal coding benchmark, Muse achieved 70.6%, while Opus 5 reached 79.4%.
The acceleration charts showcase a different ranking. Over more than 1,000 tool calls, Opus 5 exhibited the greatest improvement from the baseline (approximately 74–75%), with Muse Spark 1.2 in the middle range at around 61–69% depending on the test. Meta emphasizes that the agent continues to improve as tool call volume increases, which is desirable for a coder working over a long duration.
Notably interesting are the long-horizon and multimodal demonstrations. In stress tests, Meta indicated that Muse Code "iteratively refined GPU kernels over 1,000+ tool calls (spanning up to 24 hours) on Nvidia Hopper GPUs," showcasing its ability to enhance performance over time.
Additionally, there is a visual-coding aspect. In one demonstration, a user inputs a fly-through video of a house as an mp4 into the terminal, and Muse Code "interprets the video to create a visually appealing website with booking features." This capability of converting raw video into a functional web application aligns with the multimodal vision Meta aims to promote across the Muse series.
See the launch thread:
Here’s an example of Muse Code’s multimodal visual coding capabilities. In this demo, the user inputs a fly-through video of a home into the terminal as an mp4 file. Muse Code interprets the video and produces a visually rich website with booking capabilities. pic.twitter.com/3CAfIMYmAB
— AI at Meta (@AIatMeta) August 5, 2026
Competitive Landscape
However, Meta's entry comes later than its competitors. OpenAI's Codex already operates parallel cloud agents; DeepSeek has developed its own alternative to Claude Code, and other agentic tools like Hermes or OpenClaw already offer substantial capabilities. Muse Code’s strengths lie in its crash-safe runtime and subagent architecture rather than in benchmark excellence.
The inherent risk in agentic coding remains: an agent that can resume after a crash and continue tool calls for 24 hours is potent yet unpredictable. Meta believes developers will appreciate this level of autonomy, and it is launching the tool now.
Muse Code is available for testing through the following installation command:
curl -fsSL https://dev.meta.ai/install.sh | bash