Overview
- Multiple labs have released production-capable diffusion LLMs that use parallel denoising rather than left-to-right decoding, including Mercury 2 from Inception Labs, Gemma Diffusion from Google, Nemotron Diffusion from NVIDIA, and open-weight LLaDA.
- Diffusion LLMs generate full sequences from a noisy draft and then refine them over several steps, which lets models correct errors during sampling and use bidirectional context at every pass.
- Key algorithmic advances driving recent gains include masked diffusion (MDLM), remasking and uniform-state noise to enable error correction, block diffusion for variable lengths, encoder–decoder designs to cut repeated work, and progressive distillation to speed sampling.
- Labs report large throughput improvements and near-autoregressive quality, with common claims of 5–10× faster sampling and Mercury 2 citing roughly 1,000–1,200 tokens per second on standard GPUs, though these performance figures still need independent benchmarking.
- Diffusion has produced early scientific wins — ESM3 for proteins and NT-v3 for nucleotide design with reported wet‑lab validation — but challenges remain in matching the largest autoregressive models on absolute scale and in proving that parallel inference yields broader capability gains.