LLMs are known for requiring lots of electricity to function, and this is mostly due to how AI corps aren't (weren't?) required to make things as efficient as possible (because they're backed by essentially "infinite money cheat"). Then DeepSeek unexpectedly appeared, bringing a thing that is insanely optimized and efficient (just compare how many GPUs they needed versus how many GPUs OpenAI et al used). And since they openly released the weights, there are many sizes available as of now: from those running on their servers, all the way to models capable of being run in a modest smartphone. You read it right: these things can run without needing power-hungry GPUs, literally in the palm of the hand. It's definitely way less harm to environment, since a smartphone cannot pull more energy than it's physically capable of pulling from a USB cable (and if it's unplugged, it'll necessarily rely on the battery).

So I tried to run a distilled version of DeepSeek locally in my smartphone (my phone is not a flagship). It's not as perfect as a big model (the ones running on GPUs), but it's... Interesting. Notice the wattmeter in the notification bar: it's pulling almost a single ampere (1A × 5V = 5W) while it's "thinking". The fact I can run a damn LLM using just 5W and a mobile device is just mesmerizing.