Gemma 4 12B: The Developer Guide (developers.googleblog.com)

Gemma 4 12B has been launched as a transformative multimodal model featuring an innovative encoder-free architecture. This new release addresses significant limitations found in traditional multimodal models, which typically use separate, frozen encoders for processing vision and audio inputs. By employing a single decoder-only transformer, Gemma 4 12B enhances performance while reducing latency and memory fragmentation. The model excels in diverse tasks such as automatic speech recognition, video understanding, and coding, showcasing its versatility and efficiency. The launch of Gemma 4 12B also introduces powerful on-device developer integrations via LiteRT-LM, enabling seamless execution of local AI applications in desktop environments. For instance, the model can now function offline on Apple Silicon GPUs with applications like the Google AI Edge Gallery and Voice Edit, allowing developers to run Gemma 4 12B locally as an OpenAI-compatible API server. This integration promotes rapid development and deployment of multimodal AI applications, marking a significant leap forward for local AI capabilities.