Introducing Gemma 4 12B: a unified, encoder-free multimodal model

Google DeepMind has released Gemma 4 12B, a multimodal model that unifies image and text processing without relying on a separate encoder. Unlike traditional vision-language models that combine an image encoder with a language model, Gemma 4 directly processes pixel inputs through its transformer architecture. This encoder-free design reduces complexity and computational overhead, making the 12B-parameter model suitable for local deployment or edge devices. According to Google DeepMind, the model achieves competitive results on benchmarks like VQAv2 and COCO Captions while maintaining efficiency. For developers building AI workflows, this means they can incorporate vision-language capabilities—such as image captioning, visual question answering, or document understanding—into applications without needing to manage separate encoder components. The model is released under a permissive license, allowing fine-tuning and integration into custom pipelines. Given its size and encoder-free nature, Gemma 4 is especially relevant for resource-constrained setups or where latency matters. Developers can experiment with it via platforms that support open-weight models.