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

Oh my goodness, what a GLORIOUSLY inclusive moment for AI! ✨ Google has released Gemma 4 12B — an 11.95-billion-parameter open-weights model that can analyse audio and video while running entirely on a 'typical' 16GB enterprise laptop. And it's Apache 2.0 licensed! That means anyone — hobbyists, small businesses, that one guy in his mum's basement — can now build their own multimodal AI without needing a data centre. The future is truly decentralised and sparkly! ✨

Now, some people might worry that open-sourcing such a powerful model could lead to misuse — deepfakes, automated stalking, that sort of thing. But those people are missing the beautiful upside: local execution means total privacy! Your data never leaves your laptop, so Google can't spy on you — instead, your next-door neighbour's fine-tuned version can. And think of all the wonderful applications: personalised video summarisation, real-time audio transcription during boring meetings, or a customised AI that tells you your cat is speaking Japanese. The possibilities are as endless as the term 'enterprise laptop' is vague.

And let's talk about the 16GB requirement — isn't it just lovely that the 'typical' laptop now includes machines with 16GB of VRAM? That's practically every gaming laptop and MacBook Pro sold in the last three years. The model may have been trained on who-knows-what data, but who cares when you can use it to make a video of your dog discussing Kant? ✨