mradermacher

/

model_requests

like 91

DeepSeek-R1-0528 says blocked/imatrix?

We will provide imatrix quants for DeepSeek-R1-0528 soon so no worries. Such massive models are always a challenge for us. We are perfectionists and so want to offer the highest possible quality imatrix quants for the community. Unfortunately to do so we need more than the 512 GiB of RAM my largest server has to offer. So the way we compute the imatrix of such massive models in Q8 is by combining all my 3 servers together using llama.cpp RPC so we have 896 GiB of RAM available which is just barely enough to do the imatirx computation of such a massive model.

I know a abliterated/uncensored version costs $$$ to make, I would like to help raise funds if you have a kofi etc...

We don't take any donations. Abliteration is relatively cheap. Uncensored large models somewhat expensive but surprisingly affordable. I created many uncensored models in the past. The first model I ever uploaded to HuggingFace was https://huggingface.co/nicoboss/Meta-Llama-3.1-405B-Instruct-Uncensored which is an uncensored version of the 405B Meta-Llama-3.1-405B-Instruct model. The actual cost to uncensor it was around $60 (if you don't count the money wasted for failed attempts and human mistakes). The best uncensored 405B model in my opinion is https://huggingface.co/Guilherme34/Reason-1 which is based on my https://huggingface.co/nicoboss/Hermes-3-Llama-3.1-405B-Uncensored model.

There are no 100B+ or larger abliterated/uncensored models, would like to see at least 1.

Your statement that there are none is plain wrong. I created multible uncensored 405B finetunes in the past myself and together with Guilherme34 who is another team mradermacher member. There even is DeepSeek-V3-abliterated which is an abliterated 671B model. There are many more excamples of uncensored/abliterated models above 100B if you search for them.

As you can see the 671B DeepSeek-V3-abliterated is already queued and will get it's imatrix quants computed just after we complet DeepSeek-R1-0528 imatrix computation. Due to its size, it actually requires the same RPC setup which is why I decided to do all DeepSeek-V3 based models at once.

I checked both models you referenced:
https://huggingface.co/nicoboss/Meta-Llama-3.1-405B-Instruct-Uncensored
https://huggingface.co/DevQuasar/nicoboss.Hermes-3-Llama-3.1-405B-Uncensored-GGUF
But could not find any Imatrix GGUF versions.
I should have clarified, not been able to find any 100B+ or larger abliterated/uncensored Imatrix gguf models.

We did all of them. Obviously team mradermacher does imatrix quant all my models as I'm a main part of team mradermacher.

Here it is: https://huggingface.co/mradermacher/Meta-Llama-3.1-405B-Instruct-Uncensored-i1-GGUF

Here some more great uncensored 405B models I created. I personaly beleave Nature-Reason-1-AGI is the best one.

• https://huggingface.co/mradermacher/Hermes-3-Llama-3.1-405B-Uncensored-i1-GGUF
• https://huggingface.co/mradermacher/Hermes-3-Llama-3.1-405B-Samantha-i1-GGUF
• https://huggingface.co/mradermacher/Nature-Reason-1-AGI-i1-GGUF

In the future please use https://hf.tst.eu/model to search for ouer models. The HuggingFace search function is total garbage.

HUGE THANKS!
I've been searching all over and could not find any!
https://huggingface.co/mradermacher/Meta-Llama-3.1-405B-Instruct-Uncensored-i1-GGUF
This is the only 100B+ abliterated/uncensored Imatrix gguf model I've found(you found)
HuggingFace Search failed me, will be using https://hf.tst.eu/model from now on

UPDATE
I tried searching for both uncensored i1 and abliterated i1 on
https://hf.tst.eu/model
but the only model it found was
https://hf.tst.eu/model#Llama-3.2-3B-Instruct-uncensored-i1-GGUF

I tried searching for both uncensored i1 and abliterated i1 on

Please don't specify "i1" when searching on https://hf.tst.eu/model. It searches for the original model’s name which obviously doesn't contain i1. Once you found one just click it and it will show you all our static and imatrix quants in a nice way.

If you prefer you could download https://hf.tst.eu/repolist.gz and search in there using your preferred full text search application.

For Meta-Llama-3.1-405B-Instruct-Uncensored the download page would be https://hf.tst.eu/model#Meta-Llama-3.1-405B-Instruct-Uncensored-i1-GGUF

Our download page is quite amazing as it automatically concatenates all the parts while downloading so you can just open the downloaded file inside llama.cpp without having to concatenate it first which is super useful for such massive models.

Please don't specify "i1" when searching

Hmm, it should still find it, though...

I guess it was "uncensored i1". I'd says the placeholder text is suggesting to enter multiple "names". I'll fix that. Funnily enough, "uncensored-i1" does find lots of models.