I second unsloth models. I'm using them over blackwell-oriented nvfp4 models as they are (empirically) top quality and performance.
replyNVFP4 will be better if the model provider actually post-trained properly after quantizing.
replyWhich basically only Nvidia does, because it’s very expensive.
replyThough I’m currently working on QADing the smaller Qwen 3.5 models from FP16 teacher to NVFP4 student, to hopefully eventually apply it to 3.6 27B… harder to get right than I expected though!
deleted
reply