[MallocVoidstar]

That isn't unquantized, it's de-quantized. They went from Q5 to fp16 for use in Pytorch instead of the GGUF ecosystem.

[Taek]

I never thought people would be upscaling models by increasing quantization precision. The rationale makes sense bit its also a goofy outcome.

[nullc]

You should be able to upscale and fine tune to recover performance, I suppose!

Clearly we should train a diffusion model to denoise the weights of LLM transformer models. Yo dawg.

[throwaway9274]

Yes, that’s correct. Good correction.