Anyone else suffering from verbose, poor technical documentation and explanations with Q3.8-Flash-Next?

The model is quite good at technical work. But ask how something works, or ask it to document something, and it overkills with extremely verbose explanations, extremely dense technical language that is difficult to understand, and using strange metaphors that don't make sense. I haven't been able to reign in on it. For the record I'm using NVFP4 with FP8 kv cache (though I tried BF16 kv and I don't think it helped), so it's quite possible that full quality weights would help, but man, would be nice to find a solution. Unless Qwen4-Flash will fix this...