DiffusionGemma 改造 Gemma 4 實現每秒 1500 token，但推理能力滯後
google deepmind 發布了 diffusiongemma 的技術報告，這是一個從 gemma 4 轉換而來的文本擴散模型，在 h100 上可達到每秒 1,500 token 的速度。雖然它透過並行去噪帶來了巨大的速度提升，但在原始推理品質上仍落後於自迴歸模型。
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.