ORL Speeds Up Diffusion LMs to 2788 Tokens/s on a Single RTX 5090

Researchers from Together AI, UIUC, UC Berkeley and Stanford released ORL (On-Policy Residual Learning), a fully open approach that rewrites the speed-intelligence Pareto frontier of diffusion language models, boosting throughput 6.9x and reaching 2788 tokens/s on a single RTX 5090.

2026-10-16 ~ 2026-10-17 · 3 related posts

• ORL pushes diffusion LMs to a new speed-intelligence Pareto frontier with ~6.9x throughput gains — james_y_zou · 2026-10-16
• 2788 tokens/s on one RTX 5090: open-source ORL accelerates diffusion LLMs — arty_photography · 2026-10-17
• ORL hits 2,788 tokens/s on one RTX 5090, making diffusion LLMs faster than speculative decoding — arty_photography · 2026-10-17