Meta AI releases Muse Code and Muse Spark 1.2 for advanced coding tasks
Meta releases Muse Code, a terminal coding agent for complex software engineering tasks powered by Muse Spark 1.2, an improved model focused on code generation, debugging, and large codebase understanding. The model is trained on long-horizon tasks and demonstrates iterative self-improvement capabilities. It is available in beta for macOS and Linux with integration into the Meta Model API.
How does Muse Code work and what are its main features?
Muse Code is a terminal agent powered by Muse Spark 1.2 that handles complex software engineering tasks through planning changes, writing code, and validating results. It operates with a main agent and asynchronous background agents, uses a local event log for crash recovery, and includes built-in skills like /plan for planning, /grill for testing robustness, and /goal for achieving objectives.
How does Muse Spark 1.2 differ from its predecessor?
Muse Spark 1.2 focuses on code generation, complex debugging, codebase understanding, and end-to-end developer workflows. The model was trained on long-horizon tasks including whole-repository generation, large projects, and auto-research, and uses a self-improvement loop to better follow complex instructions.
What tasks did Muse Spark 1.2 accomplish in the kernel optimization case study?
The model iteratively optimized GPU kernels for KDA and MLA algorithms on NVIDIA Hopper GPUs through over 1000 tool calls, achieving substantial performance improvements over baseline implementations by combining kernel fusion, tiling, and algorithm-specific optimizations.
- Muse Spark 1.3 aims to compete with Opus and GPT 5.6 Sol — developer.meta.com 81 % match
- Meta releases Muse Glimmer, a small open-weight AI model for local agents — research.meta.ai 79 % match
- LFM2.5-2.6B: small and capable local AI model — liquid.ai 71 % match
- Muse Code
- Muse Spark 1.2
- Meta AI
- NVIDIA Hopper
- GPU kernels
- Meta Model API