Muse Spark 1.2 is presented as Meta’s third release in four months, continuing a rapid model-development cadence. Independent testing shows clear progress over earlier versions: at the xhigh setting, it reaches an Artificial Analysis Intelligence Index score of 54, rising from 43 and 51 for the two prior releases. The available evidence does not establish its architecture, parameter count, training scale, or training-data composition, so its lineage can only be described through this measured release progression.

The clearest practical improvement is in agentic knowledge work, where the model handles realistic assignments such as preparing presentations and analysis through an agent loop with shell and web access. Its GDPval-AA v2 result improves by 260 Elo points to 1631, while Terminal-Bench v2.1 rises from 78% to 80%, indicating stronger performance on terminal-oriented coding tasks. These gains make the model a strong fit for complex, tool-assisted research, analytical deliverables, and software workflows, although independent analysis notes that the improved performance comes with higher cost per task.