Uno claims a 3× lossless LLM speedup

A new Uno paper on arXiv adds lightweight diffusion weights to autoregressive models and reports up to 3× higher throughput while preserving the base model’s output distribution, with an 8B checkpoint outperforming 26B DiffusionGemma and Mercury 2 across its tested agentic tool-use, coding, and long-context benchmarks. The authors release code and checkpoints, so the claim is inspectable rather than a closed vendor speed chart. Builders should treat this as a serving experiment first: test exact quality and batch behavior before changing an inference stack, while Google’s DiffusionGemma release shows why parallel decoding is worth tracking.