DiffusionGemma Tech Report: Online Post-Training Pushes New Quality-Speed Pareto Frontier

bodonoghue85 · x · 2026-08-05

The tech report for DiffusionGemma is officially out. The major highlight is the introduction of online post-training, which acts as a game changer for text diffusion models.

After SFT, the team applied joint Sampler Distillation and Reinforcement Learning (SD⋅RL). This approach successfully pushes the model's quality-speed Pareto frontier into a new regime.

Related event: DiffusionGemma: Parallel Diffusion Decoding Breaks LLM Inference Speed Bottleneck(8 posts)→

Original post →

More from Research

Research channel →