Saber is a new inference engine claiming 6,645 tok/s prefill and 316 tok/s decode peak for Qwen3.8 27B. Its developer says it takes a different approach from existing engines after a month of testing.
Read it at Reddit →
← Front page