This is not a diss to Unsloth, they make great quants and really move this community forward. I've been experimenting with quanting specific sublayers based on which ones have the most outliers post Q8 quant. I basically did a BF16 to Q8_0 conversion and looked at the post quant values to compare. I found several layers that had a CRAZY high number of outliers. I'm not certain this is better, but the results are interesting! I still need to upload the Q8 quant to hugging face, but here are s...