DiffusionGemma retrofits Gemma 4 for 1,500 tokens/sec but reasoning lags
google deepmind released a technical report for diffusiongemma, a text diffusion model converted from gemma 4 that hits 1,500 tokens per second on h100s. while it offers massive speed gains via parallel denoising, it still trails autoregressive models on raw reasoning quality.
google deepmind detailed diffusiongemma, a text diffusion model retrofitted from the existing gemma 4 checkpoint rather than trained from scratch. the conversion utilizes a two-phase process combining reinforcement learning and sampler distillation to balance quality with the massive speed gains of parallel denoising. this method allows the model to generate about 1,500 tokens per second on an h100 and improve at tasks requiring bidirectional context, like sudoku or structured json output. however, the report acknowledges that quality benchmarks fall short of the autoregressive base model, partly because the architecture was not originally designed for diffusion.