When could we have "Qwen3-14B-192k-Josiefied-Uncensored-NEO-Max-GGUF"?
"Goekdeniz-Guelmez/Josiefied-Qwen3-14B-abliterated-v3" has came out for one week, so when could we have "Qwen3-14B-192k-Josiefied-Uncensored-NEO-Max-GGUF"? before "Qwen3-32B-Josiefied" came out I think that's the model in our dream~
It is qued, likely end of the week.; maybe sooner.
@DavidAU I'm glad to hear this news. However, I also want to remind you that the "Josiefied-Qwen3-14B-abliterated-v3" model is very prone to getting stuck in a loop, repeating parts of its response at the end of the conversation when approaching its current window limit, with about a 60% chance when using English and 70-80% when using Chinese.
Hmm ; might be a way to counter act this with the quant mixing.
Specific quants? Which one(s) are you using?
Neo Imatrix-MAX Qwen 14B both V2 and V3 of Josiefied-Qwen3-14B-abliterated will begin uploading in the next few hours.
RE: Repeats ; noticed this issue -> Raise temp, especially if long generation however...
Also; you may want to see this model, which uses V3 Josiefied-Qwen3-14B-abliterated in a larger merge:
https://huggingface.co/DavidAU/Qwen3-The-Xiaolong-Josiefied-Omega-Directive-22B-uncensored-abliterated-GGUF
(this is one of 3, at 22B, links on the repo page)
Thank you for still remembering me.
I am actually specifically waiting for the 128k or 192k version of that model "Neo Imatrix-MAX of Josiefied-Qwen3-14B-abliterated". Are you uploading the 128k or 192k version?
Additionally, I have already tried the new model you recommended and found that it does not follow instructions very well. For example, when asked to write a 2000-word chapter of a novel, it completes the chapter but only produces about 500 words. I don't know the original Qwen3-14B from Qwen team model can done this task well or not, but Qwen3-32B-128k from unsloth can perfectly accomplish this task, so I am very eager for the release of the "Neo Imatrix-MAX uncensored" version of that model.
RE: New model;
It may take a number of tries to reach this word count ; this is something I am noticing in Qwen models, as well as other arch types.
However, it may also be related to max context -> Qwens with higher max context levels tend to produce longer content.
According to Qwen, the 4Bs and up were trained at 128k context, but released with 40K context - this might be the root of the issue with long context generation.
All the Qwen 3 tests I did after extending the context -> almost all resulted in much longer, detailed output VS "org model at 40k".
In fact, in some cases, the output was 3-5x the requested size ; with the record at 12k.
I may release long context versions of the 22Bs ; have build/test these.
This might address the long gen issue.
That being said; larger models are better at longer gen.
RE: 14B -> releasing the 14B V3 at 192k , the V2 is failing testing.
There is a also the repeat paragraphs issues showing up ; may need to use DRY to address / fix.
I may hold back the upload until this issue can be checked.
Got it—thanks so much for your hard work and contributions to the open-source community! Really appreciate the detailed updates, and I’m very much looking forward to the release of your upcoming models, especially the long-context versions.
Thank you !'
RE: V3 14B -> Still not testing correctly.
I am going to hold off uploading and see if a I can merge it with another 14B to correct some of the issue(s).