Gemma 4 12B brings native audio to a laptop-sized model

Google released Gemma 4 12B on 3 June, adding a mid-sized model with native audio input to its open model family. The announcement says the model can run locally with 16GB of video memory or unified memory and is available under the Apache 2.0 licence.

Its architecture sends visual and audio information into the language-model backbone without the separate multimodal encoders commonly used to prepare those inputs. Google presents that design as a way to reduce memory use and latency.

Developers could obtain pre-trained and instruction-tuned checkpoints through Hugging Face and Kaggle. Google listed support from local model applications and inference frameworks, alongside cloud deployment options. The release gives teams a route to process several input types on their own hardware. Google’s comparisons with its larger models remain vendor benchmark claims; memory capacity alone does not establish performance for every device, configuration or task. Local deployment also transfers setup and hardware choices to the developer.