Qwen3.6 and gemma 4 just shown that we were still far from capability limit on small models, we still are. Huge models are good, but is ther…

Qwen3.6 and gemma 4 just shown that we were still far from capability limit on small models, we still are. Huge models are good, but is there really a point in going bigger. Just realize something. One of these "small" large language models hold most of the human knowledge. It's compressed using the best way humanity ever compressed knowledger. Quantization even helps, with smart tricks invented every few months. Ask a few gb models how to create something? it will know. Who was the president of botswana on a specific date? it will know too. Current models at 35 billion parameters and 3 billion active (that's the speed!) are beating huge 700B parameters model of a year ago. Literally. I've been honestly blown away by small models running on my laptop gpu being able to one-shot games in less than a minute. Yeah, one command, one prompt, one small game. Acceleration is now. Intelligence is a commodity and is mostly free. Well, it's only the beginning.

Source: Clem Delangue (X) | 2026-04-22