I was reviewing the comments of a video and someone said this: “That’s why it only supports the ‘native’ MXFP4 quant and not any requants from HF, including the wonderful Q2_K_S-AutoRound from Intel.” and I was interested to see if it were true.

I downloaded the model, made a modelfile for it, added the template from the native version and that’s it. worked fine first time. I haven’t tested this hard, but seems to be fine.

I am using Ollama 0.11.5 for this and not sure if older versions work with it.