Meta unveils Muse Spark 1.3 with coding gains

Muse Spark 1.3 is the latest version of the native multimodal reasoning model first introduced in April, following the establishment of Meta Superintelligence Labs. The release targets two recurring pain points in practical AI use: staying reliable over long agentic workflows and cutting wasteful back-and-forth in coding. Meta says the model now asks for clarification on vague instructions, requests help when stuck, and checks before high-impact actions, and it has been trained not to fabricate results when it fails.

The benchmark results are mixed. Muse Spark 1.3 leads on the three coding-related measures shown, including a sharp jump on DeepSWE v1.1, and it dominates the long-context MRCR range tested at 512K-1M tokens. Yet on six agentic measures, Claude Opus 5 posts the top score in four, including GDPVal-AA v2, suggesting the agentic crown is not uncontested. The company's own comparisons, which include OpenAI's GPT-5.6 Sol, are clearly promotional, though the gains over Muse Spark 1.2 are concrete.

Meta's leadership is leaning into cost and capability simultaneously. Zuckerberg's framing that frontier performance is now so cheap it does not need measuring points to an efficiency story underway. The roadmap extends beyond this release: a larger model is planned, and an open-weights version of Muse Spark 1.2 is expected after final safety fine-tuning, indicating a phased approach that may continue with a model reportedly codenamed watermelon.