Muse Spark 1.2 (xhigh)

Muse Spark 1.2 is a coding-focused large language model developed by Meta Superintelligence Labs (MSL), released in August 2026. As a successor to the Muse Spark 1.1 version, this model is specifically optimized for agentic software engineering, repository-wide understanding, and end-to-end developer workflows. The (xhigh) designation refers to the model's highest reasoning effort setting, which leverages inference-time compute scaling to solve complex technical problems that require deep planning and iterative verification.

The model is natively multimodal and features a 1-million-token context window, enabling it to process extensive codebases and documentation simultaneously. It is designed to power the Muse Code terminal agent, where it can function as an asynchronous background worker. In this role, the model continuously gathers context, validates changes, and performs automated investigations, allowing developers to delegate entire software engineering tasks—such as bug fixing across multiple files—to the AI agent harness.

Technically, Muse Spark 1.2 was trained with significantly increased compute on coding tasks compared to previous iterations, expanding its training environment to include a wider diversity of development scenarios. It features a "Contemplating" mode that orchestrates multiple subagents to optimize for long-horizon reasoning. While the model is proprietary and available via the Meta Model API, Meta has positioned it as a competitive alternative to other frontier systems, emphasizing cost-efficiency for large-scale enterprise development and complex reasoning tasks.