dailyai.report

Four times faster text generation defines DiffusionGemma, a new model from Google DeepMind. It replaces traditional autoregressive sampling with a diffusion-based approach to predict multiple tokens simultaneously. This architecture reduces latency during inference. Practitioners can now deploy high-throughput text systems without the linear bottlenecks typical of standard large language models.