DiffusionGemma 26B MoE

ORL open-sources diffusion LLM hitting 2,788 tokens/s on one RTX 5090

Researchers open-sourced ORL (On-Policy Residual Learning) and its dlmrt inference engine, which speed up diffusion language models by having them carry forward and refine their own residual predictions, plus an adjacent entropy contraction objective. At batch size 1, LLaDA-2.1-mini reaches 2,788 tokens/s on a single RTX 5090 and DiffusionGemma 26B MoE hits 7,708 tokens/s on 8×B200, with up to 6.9× higher end-to-end throughput on QwenDLM-4B. Models, training code, evaluation pipeline and engine are all open-sourced.