Gemma 4 12B Developer Guide Highlights Local Multimodal AI

Google has released Gemma 4 12B, a dense multimodal model designed for local use, with details published in a new developer guide on June 3, 2026. The company says the model uses a unified, encoderfree architecture that feeds vision, audio, and text directly into the language model backbone. The guide says Gemma 4 12B is the first mediumsized model in the Gemma family to support native audio input, and that it is sized to run on laptops with 16GB of VRAM or unified memory. Google also says a separate multitoken prediction model is available to improve local inference speed. The post outlines how the architecture removes separate vision and audio encoders to reduce latency and simplify memory use. It also describes new desktop and ondevice developer integrations, including macOS apps, a local OpenAIcompatible API server through LiteRTLM, and support for tools such as LM Studio, Ollama, Hugging Face Transformers, llama.cpp, MLX, SGLang, and vLLM.