Modlearn

Was it the chip or the software? Its headline was 10–35× lower power. Per token, though, Qualcomm was more efficient on only 3 of 12 models, and it served most models at ~6–25 tok/s.

Description read from the page itself, because the listing it was found in gave none.