Google releases Gemma 4 12B, an encoder-free multimodal model

Google DeepMind on June 3, 2026 released Gemma 4 12B, a 12B-parameter multimodal model that adopts an encoder-free (unified) architecture, removing separate vision and audio encoders, as open weights under the Apache 2.0 license. It can run locally via Hugging Face, Ollama, LM Studio and llama.cpp, operating on laptops and workstations with around 16GB of VRAM. (Google announcement, model card)