Meta Releases Muse Spark 1.3 with Enhanced Agentic Capabilities

Meta released Muse Spark 1.3 from Meta Superintelligence Labs on September 2, 2026, making the new model immediately available to users in Muse Code and the Meta Model API. [1] The company stated the update improves performance across both coding and agentic tasks, making it easier to use in real-world settings. [1] While previously available reasoning modes are accessible at launch, a "max reasoning" mode will be released after further safety testing is completed by the company. [1]

The model is designed to sustain longer-horizon work by managing multiple workflows within a single thread and collaborating directly with users. [1] Meta noted that Muse Spark 1.3 actively asks clarifying questions when prompts are ambiguous, seeks user help when stuck, and confirms before taking consequential actions. [1] Furthermore, the model adapts to user preferences during long tasks by either working silently or providing frequent updates, and it is better at preserving detailed requirements across multi-step instructions without dropping constraints. [1]

Meta’s focus on long-horizon agentic workflows suggests a broader shift toward models that act as collaborative partners rather than simple query-response engines. By implementing mechanisms to pause and ask clarifying questions or require confirmation for consequential actions, the company is directly addressing the reliability issues that often plague autonomous AI deployments in the enterprise. The delay of the "max reasoning" mode indicates ongoing caution regarding the deployment of advanced capabilities without thorough vetting. What the sources don't address: How the model's benchmark performance specifically compares to rival agentic systems like Anthropic's Claude 3.5 Sonnet or OpenAI's o1.