DiffusionGemma
2026-06-10 · Publisher : Google · Domain : Artificial Intelligence
Google DeepMind releases DiffusionGemma, an experimental 26B open model (3.8B active) that generates text by diffusion, up to four times faster than an autoregressive model on a dedicated GPU.
What changes
- Generates blocks of 256 tokens in parallel instead of one token at a time.
- More than 1000 tokens per second on an NVIDIA H100 GPU and more than 700 on an RTX 5090.
- Bidirectional attention and iterative self-correction, useful for inline editing and code infilling.
- Lower quality than Gemma 4: Google recommends Gemma 4 for production.
- The speed gain applies mainly to local inference at low concurrency.
### Specifications - Architecture: derived from Gemma 4 26B A4B MoE (25.2B total, 3.8B active) with a diffusion head. - Context: up to 256K tokens; text, image and video inputs, text output. - Apache 2.0 license; fits in 18 GB of VRAM once quantized. - Knowledge cutoff: January 2025. ### Pricing Open weights under Apache 2.0, no API pricing. ### Benchmarks | Benchmark (model card, scores published by the vendor) | DiffusionGemma | Gemma 4 26B A4B | |---|---|---| | MMLU Pro | 77.6% | 82.6% | | AIME 2026 | 69.1% | 88.3% | ### Key takeaways A research model for exploring fast local generation; not to be chosen for uses where quality comes first.
Links
Tags : LLM, Open Source, Google, Gemma, Diffusion