Meta AI Launches Muse Code Beta With Powerful Spark 1.2 Model Capable of Autonomous Coding Across Large Repositories
Summary
Meta AI launches Muse Code in beta, a terminal-based coding agent powered by the new Muse Spark 1.2 model that can autonomously plan, write, and validate code across large repositories, with benchmark tests showing it successfully optimizes GPU kernels over 1,000+ tool calls for significant performance gains on NVIDIA Hopper GPUs.
Key Points
- Meta AI releases Muse Code (beta), a terminal-based coding agent powered by the new Muse Spark 1.2 model, capable of planning, writing, and validating code across large repositories using persistent async background agents.
- Muse Spark 1.2 delivers major improvements in code generation, debugging, and long-horizon tasks through scaled-up training compute, self-improvement loops using its predecessor model, and co-training directly with Muse Code for optimized performance.
- Benchmark tests show Muse Spark 1.2 successfully optimizes GPU kernels over 1,000+ tool calls, achieving substantial performance gains on KDA and MLA kernels for NVIDIA Hopper GPUs, with the model now available via Muse Code and the Meta Model API.