Google's new open source Gemma 4 12B analyzes audio, video — and runs entirely locally on a typical 16GB enterprise laptop

Google has released Gemma 4 12B, an open-weights multimodal large language model with 11.95 billion parameters. The model is licensed under Apache 2.0 and is optimised to run entirely on a standard enterprise laptop using 16GB of VRAM or unified memory. It can process and analyse audio and video inputs in addition to text.

The release is part of Google's Gemma family of open models. The 12B parameter size makes it one of the larger openly available models that can run locally on consumer-grade hardware with sufficient graphics memory. The Apache 2.0 license permits commercial use, modification, and distribution, subject to certain restrictions. Google has not disclosed the training data composition for the model.

The model's ability to handle multimodal input locally could reduce reliance on cloud-based AI services for certain applications, potentially offering lower latency and enhanced privacy. However, the open licensing also means the model can be freely used for purposes that Google may not endorse. The company has not announced specific safety evaluations or guardrails beyond those built into the model weights.