Google drops Gemma 4 12B: tiny multimodal model runs locally on 16GB VRAM
Google has released Gemma 4 12B, a compact multimodal model capable of running locally on a laptop with just 16GB of VRAM, as the Gemma series surpasses 150 million downloads. The new weights are available on Hugging Face under an Apache 2.0 license.
Google has released Gemma 4 12B on Hugging Face under an Apache 2.0 license, according to TestingCatalog. The model was built with the same multimodal functionality as the Gemma 4 E2B and E4B variants, accepting text, audio, image, and video inputs to bring native audio and vision understanding directly to local environments without cloud dependency.
DeepMind CEO Demis Hassabis celebrated the milestone of 150 million downloads for the Gemma 4 family while announcing the new 12B parameter variant. He emphasized that the model is "incredibly powerful for such a small model" and tiny enough to run locally on a laptop equipped with just 16GB of VRAM.
The release reinforces Google's strategy of delivering capable open models that operate on consumer hardware, giving developers a locally runnable multimodal foundation model under a permissive open-source license.
The Wire · Newsletter
One careful email,
every Monday.
The week's most important AI stories, lightly edited and personally vouched for. No autoplay, no spam, easy to leave.
Comments · 0
Sign in to join the discussion.
Be the first to leave a thought.