> albeit at a lower tok/s rate

Not always! I get 80-100 tok/s from Qwen 3.6 35B-A3B on a MacBook Pro thanks to MTP. With long contexts that dips to around 50-60. However, prefill is much slower than API models. So it becomes really, really, really critical to not have cache misses.

Slower is meaningfully dumber when you’re time bounded and need all the inference time compute you can get.