Tag: inference speed

Inception’s Mercury 3 writes whole answers in parallel

The diffusion language model reports 1,068 tokens per second and a million-token context, betting speed beats left-to-right decoding.