Meta has launched Muse Spark 1.1, a significant update to its specialized coding model family designed to handle long-horizon autonomous software engineering tasks. Unlike standard models that process single files or provide basic autocomplete, Muse Spark 1.1 is architected to parse entire multi-directory repositories and propose complex refactoring strategies. Meta's technical evaluations demonstrate a 35% improvement in multi-file bug localization and dependency resolution compared to Llama 3.

The model incorporates an advanced tool-interaction loop, allowing it to execute compiler checks, run local test suites, and interpret linting errors in real-time. Meta is offering Muse Spark 1.1 under an open-weights license, enabling companies to host the model on-premise without exposing their sensitive codebases to third-party API endpoints. This release is expected to accelerate the adoption of autonomous coding agents within enterprise software teams.