Meta's Muse Wonder Full scored 57% on the Terminal-Bench 4.0 coding benchmark in Artificial Analysis's testing, up from 33% for Meta's Muse Spark 1.3