Google DeepMind unveils Gemma 4 12B, an encoder-free multimodal AI model built to run on laptops

Google DeepMind introduced Gemma 4 12B, a mid-sized multimodal model that routes vision and audio inputs directly into the LLM backbone without separate encoders. The model is designed to run locally with 16GB of VRAM or unified memory, supports native audio, uses Multi-Token Prediction to reduce latency, and is released under Apache 2.0.

Gemma 4 12B reflects a push toward local multimodal AI, where advanced reasoning and agentic workflows can run on consumer hardware rather than relying solely on cloud inference. Its open license and laptop-ready footprint could accelerate experimentation across edge devices, developer tools, and privacy-sensitive enterprise use cases.

Gemma 4 12B: A unified, encoder-free multimodal model