DiffusionGemma Tech Report: Parallel Decoding Breaks LLM Inference Speed Limits

bodonoghue85 · x · 2026-08-04

Google releases the technical report for DiffusionGemma, an experimental open-weight language model based on discrete diffusion. Unlike conventional autoregressive models that decode one token at a time, DiffusionGemma iteratively refines blocks of 256 tokens in parallel, achieving exceptionally high text generation speeds.

The model is obtained by fine-tuning the mixture-of-experts Gemma 4 model (3.8B activated and 25.2B total parameters). Its compute-efficient two-stage training pipeline uses less than 10% of the original AR model's total training token budget: the first stage employs supervised fine-tuning for bidirectional denoising, while the second combines reinforcement learning with sampler distillation to enhance generation quality.

Related event: DeepMind Releases DiffusionGemma for Fast Text Generation(5 posts)→

Original post →

More from Models

Models channel →