Google open-sources 26B text-diffusion MoE; Pichai: generation speed like a racehorse

谷歌开源26B文本扩散MoE，劈柴：生成速度像赛马一样快

Google open-sourced DiffusionGemma, a 26B MoE model that activates only 3.8B parameters at inference. Instead of generating tokens one by one, it drafts 256-token blocks in parallel, hitting 1,000+ tokens/sec on an H100—up to 4× faster than autoregressive models. Output quality is lower than standard Gemma 4, so Google still recommends the autoregressive version for production. It ships under Apache 2.0, fits quantized on consumer GPUs with 18GB VRAM, and targets latency-sensitive nonlinear tasks like inline editing and code completion.

Why it matters: Google open-sourced a 26B text diffusion model that skips autoregressive decoding, activating only 3.8B params at inference and hitting 1,000+ tok/s on a single H100. Apache 2.0, with concrete speed comparisons and mechanism details — directly useful for inference folks. Not s...