Image default is beam 5 on x86 (1 on ARM). Benchmarked on framework-pt (i5-1135G7, base-int8, real speech): beam 1 transcribes identical text ~45% faster; extra CPU threads made it slower, so only the beam changes. App revision 3.4.2 > image 3.4.1 so catalog-driven nodes roll the args change; a 3.4.1-1 suffix would semver-compare LOWER and never ship. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>