Google DeepMind Launches Gemma 4 12B, a Multimodal AI Model Built to Run on Consumer Laptops

Google DeepMind launches Gemma 4 12B, a multimodal AI model that runs locally on consumer laptops with just 16GB of VRAM, featuring a groundbreaking encoder-free architecture that processes vision and audio directly through its LLM backbone for lower latency and memory use, available now on Hugging Face and Kaggle under an Apache 2.0 license.

• Google DeepMind is launching Gemma 4 12B, a new multimodal AI model designed to run locally on consumer laptops with just 16GB of VRAM, bridging the gap between the smaller E4B and the larger 26B MoE model.
• Gemma 4 12B introduces a novel encoder-free architecture that processes vision and audio inputs directly through the LLM backbone, eliminating traditional separate encoders to reduce latency and memory usage.
• Released under an Apache 2.0 license, the model is available now on Hugging Face and Kaggle, supports popular developer tools like Ollama and LM Studio, and comes with Multi-Token Prediction drafters to further reduce inference latency.