Just kidding. Are there any distills that actually improve a model's quality? I remember the Qwen R1 8B distill improved the model, but since then, I don't remember ever using a distilled model that was better than the base model. Unless Mythos (or GPT-5.6) is some magical model where only a couple hundred samples will make Qwen-3.6, Qwen-3.5, and Gemma-4 models better I don't care about them. What happened to good distills and why do people only use 250 samples now? submitted by ...