Meta AI releases Muse Spark 1.3 with improved agentic capabilities and coding efficiency

Meta AI Research released Muse Spark 1.3, built on feedback from large-scale use of Muse Code and Meta Model API. Five agentic improvements include: long-horizon tasks with multi-threading where the model builds context from conflicting sources and fills plan gaps; proactive collaboration including asking clarification questions and confirming before consequential actions; complex long-instruction following with fewer dropped constraints; multi-task routing to correctly map prompts to tasks in messy single-thread contexts; and self-awareness calibration to recognize capability boundaries instead of hallucinating. For coding, internal engineer comparisons showed approximately 20% fewer tool calls and 25% less token consumption, with cleaner code style and fewer unnecessary turns.