Top AI Product

Google DeepMind just open-sourced DiffusionGemma, an experimental language model that doesn’t write text one token at a time. It starts with a block of 256 noise tokens and iteratively refines them in parallel — about 20 tokens per forward pass, roughly 1,500 tokens/s on a single H100. That’s faster than autoregressive models even with speculative decoding turned on.

Why this one is different

Every serious diffusion LLM before this was trained from scratch. DiffusionGemma is fine-tuned from Gemma 4 MoE (25.2B total, 3.8B active) using under 10% of the original training token budget. Translation: any strong AR model can become a diffusion model cheaply. It keeps thinking mode, multimodal input, and long context — and can still fall back to plain autoregressive generation, which points straight at hybrid diffusion-AR decoding.

Open weights, ready to run

The weights are on Hugging Face as google/diffusiongemma-26B-A4B-it, Apache 2.0. Obvious use cases: latency-critical agents, batch code generation, anything where decoding speed is the bottleneck. The technical report pulled 122 points on HackerNews — the most technically substantial release of the day.

You Might Also Like

• Google Ships Gemma 4 Multi Token Prediction Drafters 2 7 3 5x Faster Inference Free
• Google Deepmind Aletheia Just Solved Math Problems Nobody Could Heres why That Matters
• Ggml Llama cpp Joins Hugging Face and Honestly it was Only a Matter of Time
• Gemini Canvas in ai Mode Google Just Turned Search Into a Creative Workspace
• Google ai Edge Eloquent Lands on ios With Zero Fanfare Free Offline Dictation Powered by Gemma