Muse Spark 1.1 (xhigh) is a multimodal reasoning model developed by Meta Superintelligence Labs, released in July 2026 as a significant upgrade to the Muse Spark series. Designed primarily for agentic workflows, the model is optimized for autonomous planning, tool use, and complex software interaction. The "xhigh" designation corresponds to the maximum reasoning effort setting, which scales inference-time compute to maximize performance on intensive scientific reasoning, enterprise coding, and knowledge-retrieval tasks.
The model features a 1-million-token context window supported by structural capabilities for active context management and compaction, allowing it to maintain coherence over long-horizon projects. It is architecturally distinguished by an Action Reasoning Transformer layer, which facilitates a "plan-act-observe-adjust" loop. This enables the model to orchestrate multi-agent systems, where it can act as a primary agent delegating specific subtasks to parallel subagents to optimize end-to-end latency.
In addition to advanced text and image perception, Muse Spark 1.1 is trained for native computer use, allowing it to automate workflows across multiple desktop applications. It can autonomously generate automation scripts or interact directly with user interfaces when text-based scripting is unavailable. On industry benchmarks, the model has demonstrated significant performance in coding accuracy, specifically in diagnosing complex bugs and executing large-scale code migrations within massive enterprise repositories.
Muse Spark 1.1 is accessible via the Meta Model API, marking a shift toward proprietary, paid model access for Meta's high-end reasoning systems. When prompting the model, developers are encouraged to use developer messages rather than traditional system messages to define tone and standing rules. The model also supports stateful interactions through a specialized API that tracks session history via a unique response identifier to improve multi-turn reliability.