I checked how it performs in long run (prediction) on 4 x Raspberry Pi 5:

* pos=0 => P 138 ms S 864 kB R 1191 kB Connect

* pos=2000 => P 215 ms S 864 kB R 1191 kB .

* pos=4000 => P 256 ms S 864 kB R 1191 kB manager

* pos=6000 => P 335 ms S 864 kB R 1191 kB the

Technically, should be enought 2 machines with 128 GB RAM or 4 machines with 64 GB RAM (Q40 weights).

For example https://github.com/nocode-js/sequential-workflow-designer is designed in this way, that you program some logic by setting predefined actions in a proper order. You program at the high level.