How it adds up

Stack the five contributions and you get the spec sheet for Mercury, Inception’s commercial model family and, by their description, the first commercially available diffusion LLM: MDLM’s training recipe, Block Diffusion’s variable-length caching, ReMDM’s self-correction, guidance for control, and d1’s reasoning recipe — running on FlashAttention-class kernels, aligned with DPO-class methods, chasing DDIM’s goal of ever-fewer steps. Inception claims roughly 5× the speed of comparable AR models; whether dLLMs reach frontier quality is still an open bet, but you now know exactly which obstacles have been cleared, which papers cleared them, and with which equations.

← Group 2: the diffusion-LLM line The complete reference list →