Google Releases Gemma 4 12B With an Encoder-Free Multimodal Architecture
Google DeepMind launched Gemma 4 12B on June 3, 2026. The 12B model processes text, images, and audio directly in the language model backbone without separate vision or audio encoders. It can run in 16GB of VRAM, uses a multi-token prediction drafter to reduce inference latency, and reportedly approaches the 26B MoE on benchmarks. The weights are available under Apache 2.0 on Hugging Face and Kaggle.
Visit original source