# Google Releases DiffusionGemma: 26B Diffusion-Based Model Delivers 4x Faster Local Inference

_Tool · published 2026-06-11_

Google has introduced DiffusionGemma, an experimental open-source 26B Mixture of Experts language model that replaces token-by-token autoregressive generation with parallel text diffusion, yielding up to 4x faster inference on dedicated GPUs. Released under an Apache 2.0 license, the model activates only 3.8B parameters at runtime, fitting within 18GB VRAM when quantized — putting it within reach of high-end consumer hardware such as the NVIDIA GeForce RTX 5090 and NVIDIA H100.

The architecture generates 256 tokens simultaneously using bidirectional attention, making it particularly suited to non-linear tasks like code infilling, inline editing, and real-time text manipulation — use cases where standard autoregressive models face latency bottlenecks. Google notes that output quality is lower than its production Gemma 4 family, positioning DiffusionGemma as a research and developer tool for speed-critical local workflows rather than a replacement for high-fidelity deployments. Fine-tuning partner Unsloth has already demonstrated the model solving Sudoku, a task that exploits its bidirectional attention advantage.

## Sources
- [blog.google](https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/)
---
Canonical: https://genbuzz.news/posts/google-releases-diffusiongemma-26b-diffusion-based-model-delivers-4x-faster-local-inference