tested qwen 3.8 flash next IQ3_XXS for a week. the 3-bit version supposedly passes more benchmarks, but the reasoning is wobbly, filled with unsettling oddities, and the mistakes made are more surprising. so i'm going back to IQ2_S, which i was already content with.
0 Boosts
0 Likes
@lritter Honestly, I expect people to find a model or a setup that's basically "good enough for most practical applications" and then just use it.