Meta launches Muse Code, an AI agent for working with large code bases

Muse Code is currently in beta testing, TechCrunch reports . It is capable of performing full-fledged software development tasks on large repositories: planning changes, writing code, and verifying results.

The tool, which can be installed with just one command, is based on the latest Muse Spark 1.2. It can handle large projects by running its own agents that run concurrently.

Meta tested Muse Spark 1.2’s ability to thermatically optimize GPU kernels over 1,000 tool calls (up to 24 hours). Using the Muse Code agent-based coding environment, the model independently writes, compiles, profiles, and consistently improves kernel performance over the baseline implementation. Testing was conducted on KDA and MLA kernels for NVIDIA Hopper GPUs, and the AI ​​agent demonstrated significant performance gains over baseline.

“When a task is large enough, it’s split between separate subagents that run in parallel in isolated work trees,” Zuckerberg explained of Muse Code. “Your working copy remains intact. In testing, the tool simultaneously created six features for the game without any conflicts.”

Meta's Head of Artificial Intelligence and Head of Meta Superintelligence Labs Alexander Wang noted that Muse Code can be an extremely successful solution for many workflows and usage scenarios, primarily considering its cost.

Meta to start producing its own AI chip in September to reduce dependence on Nvidia and AMD

On the topic

Meta to start producing its own AI chip in September to reduce dependence on Nvidia and AMD

Meta introduces Muse Spark 1.1 multimodal model for AI agent development

On the topic

Meta introduces Muse Spark 1.1 multimodal model for AI agent development

Read the country's main IT news in our Telegram

On the topic

Read the country's main IT news in our Telegram