Post #3121840

There's also speculative decoding and adjacent techniques getting traction, increasing performance of the models on the same hardware.