Mercury Refreshed

Inception describes its refreshed Mercury as a scaled-up version of the same diffusion approach to text. Rather than predicting words one at a time, a diffusion language model starts with a rough sketch of a whole block of text and refines it over several passes, so many positions change together. Inception's refresh announcement says the new version uses larger models and more training data, an upgraded denoiser (the part that performs each refinement pass), and new training objectives, inference algorithms and serving engine. The company reports improvements in coding, instruction following, math and knowledge recall. Those are Inception's own benchmark claims, and the comparative speed and price statements in the post were not checked here. Earlier coverage of the original Mercury release gives the background.