Meta Launches Muse Code and Muse Spark 1.2 for Advanced Coding
Meta has unveiled Muse Code alongside Muse Spark 1.2, an upgrade to its Muse Spark model that focuses on coding capabilities, particularly in long-sequence agentic tool utilization. This update showcases enhancements in code generation, intricate debugging, understanding codebases, and comprehensive developer workflows. Muse Spark 1.2 has significantly increased training compute for coding tasks and broadened the diversity of the training environment, while retaining its strengths in general agent performance. Co-trained with Muse Code, the model incorporates the Muse Code toolset to optimize compatibility, utilizing rejection sampled harness trajectories and recipe enhancements for objectives, compaction, and subagents. The update is supported by extensive training on long-horizon coding tasks, including entire repository generation and large-scale projects. Additionally, a comparison of SVG pelican images generated by Muse Spark 1.1 (dated July 9th) and 1.2 illustrates a small yet significant improvement. This information was shared by Simon Willison on August 5, 2026, in a link post that also offers a sponsorship option for a curated email digest on significant LLM advancements.
Key facts
- Meta released Muse Code and Muse Spark 1.2.
- Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1.
- Improvements include code generation, complex debugging, codebase understanding, and end-to-end developer workflows.
- Training compute on coding tasks was significantly scaled up.
- Training environment diversity was expanded.
- Muse Spark 1.2 was co-trained with Muse Code.
- Training included rejection sampled harness trajectories and recipe optimizations.
- The model was trained on long-horizon coding tasks like whole-repository generation.
- Simon Willison posted the announcement on August 5, 2026.
Entities
Institutions
- Meta
- Simon Willison