Hacker News Reader

I find the qwq 32B a bit like that. I asked for a recipe for something in minecraft 1.8, and it was page after page of 'hmm, that still doesnt look right, maybe if I try...' although to be fair I did ask for an ascii art diagram for the result. It will be interesting to try a DeepSeek 32B qwq if that is planned, because otherwise pretty happy with it.

I just wish that less development chat was happening within walled gardens because none of these seem to be much help with Zig.

I noticed the smaller the model (be it quant or parameters as the cause) the faster it'd run.... but the longer it'd fight itself. For the same Calc II level problem all models were able to eventually get an answer but the distilled Qwen-32B at Q6 quant was fastest to actual answer completion.

That's the point.. the rambling is their reasoning process.

I think you’ve missed the point - the rambling is not desired, thinking could be improved if, like another commenter suggested - there was a length penalty applied.

You want thinking, but you want to penalise rambling, for many, many reasons.

They need to be trained with a small length penalty