Google launches DiffusionGemma, an open model delivering up to 4x faster text generation via diffusion-based decoding
Category: AI Models • Open Source
What it is
Google released DiffusionGemma, an experimental 26B MoE model (Apache 2.0)
that generates entire blocks of text in parallel instead of token-by-token. It
delivers up to 4x faster inference – 1000+ tokens/sec on an NVIDIA H100, 700+
on an RTX 5090 – while activating only 3.8B parameters, fitting within 18GB VRAM
when quantized. Output quality trails standard Gemma 4, so Google
recommends it for speed-critical local workflows like in-line editing and code
infilling, not production-grade output.
Why it Matters for Enterprises
A practical option for low-latency, on-device AI features - interactive tools, IDE
plugins, rapid prototyping - without cloud round-trips. Not a replacement for
production LLMs, but a strong fit for speed-sensitive internal tooling.
