mradermacher

/

model_requests

like 71

Sure can, although it might be delayed one or two days because of the two 700B deepseek models in the queue. You can check how it progresses at http://hf.tst.eu/status.html

However, I see bartowski has quants of this model, which are likely imatrix ones. His training data is a subset of ours (about half of it), so the quants should be very comparable, and you might want to have a look at them in the meantime.

Thank you, I will try bartowskis for now

@sm54 First imatrix quants of FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview are now avilable under https://huggingface.co/mradermacher/FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview-i1-GGUF/tree/main - the remaining imatrix quants will appear within the next few hours. Sorry for the delay caused due to DeepSeek-R1.

His training data is a subset of ours (about half of it), so the quants should be very comparable, and you might want to have a look at them in the meantime.

I once compared them and the quality difference is quite minor (except for some MoE models where his dataset doesn't cover all experts) despite us training on a twice as big imatrix dataset but in the end our imatrix quants are always slightly better as we train them on the same dataset as bartowski with the addition of a lot of high quality proprietary data so if you like the model and intend on using it long-term I recommend switching to our imatrix quants.

@sm54 First imatrix quants of FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview are now avilable under https://huggingface.co/mradermacher/FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview-i1-GGUF/tree/main - the remaining imatrix quants will appear within the next few hours. Sorry for the delay caused due to DeepSeek-R1.

His training data is a subset of ours (about half of it), so the quants should be very comparable, and you might want to have a look at them in the meantime.

I once compared them and the quality difference is quite minor (except for some MoE models where his dataset doesn't cover all experts) despite us training on a twice as big imatrix dataset but in the end our imatrix quants are always slightly better as we train them on the same dataset as bartowski with the addition of a lot of high quality proprietary data so if you like the model and intend on using it long-term I recommend switching to our imatrix quants.

I've always found the mradermacher quants to work better than bartowskis, and given what you've said that makes sense. I find the bartowski quants to be more prone to errors for coding tasks compared to the mradermacher ones.