# ORL speeds up diffusion language models up to 7×

> The new ORL framework lets 4B diffusion models decode faster while keeping quality, reaching over 5,000 tokens per second on a single GPU.

Oossa · 2026-10-16 · https://oossa.com/en/orl-speeds-up-diffusion-language-models-up-to-7

Researchers led by Yifan Yu introduced ORL, an on‑policy residual learning method for diffusion language models. ORL trains the model on its own prediction paths and adds an entropy‑contraction loss to sharpen outputs. On a 4 B model converted from Qwen3‑4B, ORL boosted tokens per forward by up to 6.2× and overall throughput by about 6.9×, matching or beating quality of autoregressive specs. The same technique gave a 1.5× speed gain on LLaDA‑2.1‑mini and modest gains on DiffusionGemma.

## The facts

- Paper published on 2026-10-16 ("PUBLISHED: Fri Oct 16 2026").
- ORL achieved up to 5,046 tokens/s on a single B200 GPU.

## Why it matters

Developers can deploy faster, high‑quality text generation on consumer‑grade GPUs, reducing cost and latency for applications like chatbots.

## Sources & references

1. [On-Policy Residual Learning for Diffusion Language Models](https://arxiv.org/abs/2610.20687) – arXiv cs.LG, 2026-10-16

Last updated: 2026-10-16