Google releases Gemma 4 12B, a multimodal model without separate encoder

Google's new Gemma 4 12B handles text and images in a single unified architecture, eliminating the separate encoder-decoder pipeline. For indie developers working with vision-language tasks on limited hardware, this consolidation could mean faster inference and simpler deployment without sacrificing capability.