Using something like Taalas' hardcoded model as opposed to running one on general purpose GPUs, flexible but power-hungry.

https://www.cnx-software.com/2026/02/22/taalas-hc1-hardwired...