Text generation just got 4x faster — and it's open source
On June 10, 2026, Google DeepMind dropped a bombshell: DiffusionGemma, a 26-billion-parameter Mixture-of-Experts (MoE) model that ditches the traditional autoregressive approach to text generation in favor of — you guessed it — diffusion.
How is this different?
Every LLM you've used so far (GPT, Claude, Llama, Gemini) generates text one token at a time, left to right. It's sequential. Predictably, it's also slow, especially for long outputs.
DiffusionGemma flips the script. Borrowing from how image generators like Stable Diffusion and DALL·E work, it starts with pure noise (random tokens) and iteratively denoises the entire sequence in parallel. The result? Text is generated up to 4x faster than comparable autoregressive models — with competitive quality.
Open source and practical
Under the Apache 2.0 license, DiffusionGemma is fully open-weight. The 26B MoE architecture means only a subset of parameters activates per token, keeping inference efficient even on consumer hardware. NVIDIA has already partnered with DeepMind to optimize it for local RTX GPU inference via the RTX AI Garage.
Performance highlights
- ~4x throughput vs. Gemma 4 27B at comparable quality
- 26B MoE — efficient sparse activation
- Supports 8k+ context out of the box
- Built on top of the Gemma 4 architecture
What this means
This is the first serious attempt to bring diffusion to language at scale, and it's open. If this catches on, the era of painfully sequential text generation might finally be ending. Real-time AI conversations with zero perceptible lag? That's suddenly a lot closer.
Try it yourself — weights are live on Hugging Face and Kaggle, and NVIDIA's optimized local builds are available now.
Tags: ai, opensource, machinelearning, googledeepmind
Top comments (0)