Google Expands Gemma 4 With 12B Unified Model and Faster On-Device Inference Tools

Google DeepMind has broadened its open-weight Gemma 4 family with a new encoder-free "12B Unified" multimodal model and added Multi-Token Prediction (MTP) and Quantization-Aware Training (QAT) checkpoints, giving developers faster, lower-memory deployment options that run from edge devices to workstations under the permissive Apache 2.0 license.