

Not true. 500Mb models suck ass and are just here for fun. A lot of local models in the 2.5Gb range can run on a phone and produce very coherent output on par with free-to-use LLMs without actually destroying the planet (while using them I mean, training is still a nightmare).
“Fun” fact, political bias is baked in the local models too, don’t ask Qwen3 about what happened in Tiananmen Square in 1989…




And guess what, the company I work in proudly announced today they will replace our internal chatbot with ChatGpt Enterprise. It’s like they gladly and actively increase their dependency on an enemy state.