Introduction

In the previous post I covered keeping a home RTX 5090 (Blackwell) alive 24/7 against the GSP hang. This time it's about the model that runs on top of it — a record of standing up Google's diffusion language model DiffusionGemma locally on that 5090.

Let me lead with the conclusion. It's genuinely fast — roughly ~300 tok/s by feel. But the surrounding ecosystem is still immature (Open WebUI's web search doesn't integrate, among other things), so it earned a place as an "experiment to test the speed" rather than a daily driver . So this post isn't a victory lap about landing the best local model — it's a record of the one-time pioneer cost you pay when you run a bleeding-edge model on a bleeding-edge GPU .

Why not ollama, why vLLM

DiffusionGemma isn't a normal autoregressive model. It generates via block diffusion — autoregressive between blocks, parallel diffusion within a block — and that's the source of the speed.…