Google Unveils Gemma 4 12B: Democratic Multimodal AI for Consumer Hardware

Google has released Gemma 4 12B, an open-source 12-billion parameter multimodal model that processes text, images, video, and audio natively using an encoder-free architecture. The report highlights its near-parity with the larger Gemma 26B model while running on as little as 16GB of memory, describes architectural innovations (lightweight visual embeddings, direct audio projections), and notes distribution and fine-tuning options via Ollama, HuggingFace, Kaggle, and Unsloth under an Apache 2.0 license.