DiffusionGemma Is Fast Because It Stops Pretending Text Has to Be Written Left to Right

Short summary

Google DeepMind released DiffusionGemma, an open-weight text diffusion model achieving ~1,500 tokens/sec on a single H100 versus ~303 for autoregressive Gemma 4. It denoises 256-token blocks in ~12 steps instead of generating left-to-right. The model trades raw capability for speed—scoring lower on AIME, LiveCodeBench, and GPQA—but wins on tail latency for low-concurrency agent workloads. The practical takeaway is hybrid routing, not replacement of autoregressive models.

• •DiffusionGemma hits ~1,500 tok/s vs 303 tok/s for AR Gemma 4 on H100
• •Lower benchmark scores than AR baseline—faster, not smarter
• •Best fit: low-concurrency, latency-sensitive agent loops; AR still wins at high batch

Generated with AI, which can make mistakes.