With oMLX 0.7.4.dev1, a second Mac can according to the release notes support processing of long prompts while the large model stays on the main device. This is explicitly a development version, not a stable release. The published benchmarks come from the project itself.
Source: jundot on GitHub ↗ via @CerebralCoding_
Context: Single source: GitHub release; development version; benchmarks from the project itself.