Google launches DiffusionGemma, an open-source model that's 4x faster on local devices

Google has launched DiffusionGemma, an open AI model based on diffusion technology. It's 4 times faster for local reasoning than traditional autoregressive models. Instead of generating text one word at a time, it works on all words together, making it much quicker on local devices with limited bandwidth.

The model is open-source under Apache 2.0 license and available on Hugging Face. It can fix mistakes as it generates, reaching 1479 tokens per second with just 0.84 seconds overhead. It scores 30.9% on LiveCodeBench and 89.6% on HumanEval, matching Gemini 2.0 Flash-Lite. Math performance shines with 23.3% on AIME 2025, beating the comparison model's 20.0%. However, it lags in scientific reasoning (40.4% on GPQA Diamond vs 56.5%). On a single H100 GPU, it generates 1000 tokens per second; on DGX Station, up to 2000 tokens per second.