
Google DeepMind’s Gemma 4 Technical Report introduces a new generation of open-weight AI models designed for different computing needs. The family includes E2B, E4B, 12B, 26B A4B, and 31B variants, spanning Dense and Mixture-of-Experts architectures. All models support text and image inputs, while E2B, E4B, and 12B also support audio. The 12B model features a unified architecture that processes images and audio without separate encoders.
A configurable thinking mode supports reasoning tasks such as mathematics and coding. Context windows extend to 128K tokens for smaller models and 256K for larger variants, enabling them to process lengthy inputs. Attention optimizations, quantization, and multi-token prediction enhance memory efficiency and inference speed. The report also outlines pre-training methods, instruction tuning, performance evaluations, and safety measures. Released under the Apache 2.0 license, Gemma 4 enables researchers and developers to customize and deploy models for their own applications.