Request BF16.GGUF
"Dear mradermacher team, thank you so much for converting this model to GGUF. I really like this model, and I would also like to request a BF16 GGUF version. Thank you, mradermacher."
That's one of the craziest requests we got for a while as the original GPT-OSS-20B model was trained in 4 bits per wight. The only reason the finetuned model was uploaded in BF16 is because whoever created it lacked the hardware required to finetune it in MXFP4 and so upscaled it to BF16. We can provide a BF16 GGUF if this is really something you want but realistically everything above Q4 will not make any difference. Are you sure that this is something you want because creating such a GGUF seems like a waste of resources?
You make a valid point regarding the efficiency. That said, I prefer to have a GGUF that matches the current BF16 weights exactly without any further quantization logic applied, even if the benefit is negligible. If you could upload the BF16 version, I would be very grateful. I like this model.