Smaller models are getting scary good.

I am still processing this lol. I had Gemini 3 Pro Deepthink try to solve a complex security puzzle (which was secretly an unwinnable paradox). It spit out this incredibly professional-looking, highly structured answer after about 15 minutes of reasoning. Just for fun, I passed its solution over to Gemma 4 (31B) (with tools enabled). Gemma completely tore it apart. It caught a hard physical constraint violation and a fake math equation that Gemini tried to sneak by me to force the answer.