Meta Launches Muse Spark 1.3 to Curb Agent Hallucinations

Meta Superintelligence Labs has launched Muse Spark 1.3, an incremental update to its agentic coding model aimed at resolving common production failures. In internal testing against its predecessor, Muse Spark 1.2, the new version reduces tool calls by approximately 20 percent and consumes about 25 percent fewer tokens. These efficiency gains directly lower the operational costs and latency associated with running long-horizon agentic workflows.

The update focuses heavily on calibration and reliability. Instead of confidently fabricating answers when encountering difficult tasks, Muse Spark 1.3 is trained to recognize its own limitations. It now asks clarifying questions when prompts are ambiguous, flags when it gets stuck, and requires confirmation before executing irreversible actions. This behavioral shift prevents agents from drifting away from initial instructions or making destructive errors during multi-step operations.

Practitioners can use the model to sustain open-ended work across multiple simultaneous workflows within a single thread. The model can synthesize context from messy sources, such as drafting a computational fluid dynamics report from raw simulation data and a STEP file. Meta achieved these improvements by training the model across diverse agentic harnesses, incorporating more long-horizon coding tasks, and strengthening its defenses against prompt injection.

Muse Spark 1.3 is currently available through Muse Code and the Meta Model API at dev.meta.ai under existing API pricing. While its highest reasoning tier, max reasoning, is currently delayed pending safety evaluations, developers can access other reasoning modes immediately. Looking ahead, Meta plans to release larger models and publish the open weights for Muse Spark 1.2, providing the developer community with a customizable base for agentic fine-tuning.