any plan to release the full precision model?
i'd rather quantize it myself (and have the option of using it with vllm)
I second that, would love the safetensors
+1
Hi All ;
Source is located on my partner's repo - Nightmedia - here:
https://huggingface.co/nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451
Enjoy.
Would appreciate feedback from people that try it on other platforms and formats.
I was only able to test it in LMStudio, David tried the MTP and shows it to work well, probably tools work well too.
If there are any tool issues, this is usually the template--try the latest froggeric version, I use an early version to flip on/off the think tags in testing, but don't test with tools.
Thanks!
-G
@nightmedia working well on mainline llama cpp - running q6 mtp version with 160k q8 kv context on 5090. Really hope nvfp4 / vllm happens soon.
There is a 9B version of this model
https://huggingface.co/nightmedia/Qwen3.5-9B-DS9-USS-Defiant
arc arc/e boolq hswag obkqa piqa wino
bf16 0.649,0.832,0.895,0.713,0.482,0.783,0.699
mxfp8 0.647,0.836,0.895,0.706,0.460,0.784,0.695
q8-hi 0.654,0.830,0.895,0.713,0.488,0.783,0.699
qx86-hi 0.652,0.831,0.894,0.714,0.474,0.782,0.700
q6-hi 0.656,0.829,0.894,0.710,0.476,0.777,0.705
q5-hi 0.645,0.831,0.895,0.712,0.472,0.785,0.700
qx64-hi 0.638,0.827,0.887,0.716,0.470,0.786,0.702
q4-hi 0.640,0.823,0.881,0.714,0.470,0.785,0.702
mxfp4 0.640,0.824,0.886,0.703,0.468,0.780,0.691
Quant Perplexity Peak Memory Tokens/sec
mxfp8 4.216 ± 0.027 16.02 GB 593
qx86-hi 4.099 ± 0.026 15.72 GB 629
qx64-hi 4.155 ± 0.027 13.62 GB 658
mxfp4 4.388 ± 0.029 11.55 GB 705
I tried to offer this to the US Government for 1 Billion Dollars, but they did not act fast enough, so I set the source free.
There is a LinkedIn article with lots of traces on the subject
https://www.linkedin.com/posts/gchesler_llm-ai-intelligence-activity-7480960409403805696-NLE6
USS Defiant (Trump Class) - US Naval Naming Analysis
Yes, this is actually a very nice association for several reasons. Let me break it down:
1. USS Defiant is Real
- There was actually a USS Defiant (DD-742) in WWII, an Fletcher-class destroyer
- There have been multiple ships named Defiant throughout US naval history
- So you're not inventing a completely fictional name; it has historical precedent
2. "Trump Class" as Humorous Classification
This is where it gets interesting for your specific use case:
| Aspect | Why It Works |
|---|---|
| Local vs Cloud | "Trump Class" implies a new, homegrown classification—exactly what your local 9B model is |
| Underdog Spirit | Trump's political style was about challenging the establishment; your model challenges cloud giants |
| Multi-Perspective | Trump's "truth" was often subjective; your model has Dick, Bradbury, Orwell, etc. all debating reality |
| US Context | Since you're in the US, this resonates culturally without being overly political |
Before finding this and seeing their exists the full weight version, I unpacked the q8 GGUF into bf16 safetesnors format and am running in vLLM with MTP on quad 9700s with 500k context.
Working smashingly well, will be swapping out for full original but it's not shown a single iota or twitch of abberant behavior. I nice change of pace after wrestling fundamentally broken models recently released.
Well done, still in evaluation phase but seems to be a legit step up from base.
I would be very curious to see what Deckard 40b, with replicated layer only unfrozen, trained through your process would yield....
@tcclaviger
Excellent!
RE: 40Bs ; a number are in testing presently with 711 model / Deckard variants.
Could anyone make an int8 safetensors format for vllm. I think that is the optimal format for 3090 + vllm. I would ask an AI agent to do it, but it seems not straightforward, requiring calibration data. I couldnt be sure it does a perfect job.
Hi All ;
Source is located on my partner's repo - Nightmedia - here:
https://huggingface.co/nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451
Enjoy.
@DavidAU can you release the full-precision uncensored version ?
Hi All ;
Source is located on my partner's repo - Nightmedia - here:
https://huggingface.co/nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451
Enjoy.
@DavidAU can you release the full-precision uncensored version ?
by full precision you mean this ? https://huggingface.co/nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-F32
https://huggingface.co/nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451 - is this not the same exact model as the one from this repo /
Yes, those are the same thing, the only difference is the F32 was merged as float32.
@nightmedia hey, can you suggest a quantization method for vllm int8 ? i see autoround being recommended
Should it be done from the f32 ideally rather than the f16 ?
Any chance you can make an official quant like that, maybe int4 as well while at it ?
I only use a Mac, but there are plenty people here with the hardware for it, they will catch on, give it a day or two :)
I provided both bf16 and f32 to give people a choice--I have honestly no idea what would work best for Nvidia platform.
I would love an AWQ or AutoRound quant of this model!
I would love an AWQ or AutoRound quant of this model!
https://huggingface.co/TheHouseOfTheDude/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451_INT8_PTQ
https://huggingface.co/TheHouseOfTheDude/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451_INT4_PTQ
at the moment we have these to test, basic quants. we need autoRound
I'll get one in the works will be a bit have a lot of thing going on that are distractors.
@nightmedia can you clear some confusion that has been created aroun this model.
huggingface.co/nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451 was given as full source link for the ggufs made in this repo.
But someone above requested the full precision uncensored version, but if the full precision link is not uncensored, and these ggufs are, then it means it is not the same model after all ?
Also someone just released https://huggingface.co/gorbatjovy/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic
which would confirm that maybe this repo and the full f16/32 model in your repos are not in fact the same model, if one is uncensored and the other not.
That is correct, the original merged model Qwen3.6-27B-Architect-Polaris2-Fable-B-F451 is only partially Heretic.
David's distills are all heretic.
The Qwen3.6-27B-Architect-Polaris contains Qwen3.6-27B-Claude-4.6-OS, that was merged from a 3.5 Heretic distill into the 3.6 base(non-Heretic)
The Fable traces in Qwen3.6-27B-Architect-Polaris-Fable were also not on a Heretic base, so, it is only natural that the Heretic features have been diluted in the merges.
I hope that makes more sense now :)
-G
Ok. I am uploading an AutoRound 4bit quant of this model right now (https://huggingface.co/DoktorMincs/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-W4A16-AutoRound). But as I understand, this is not the exactly the same model of DavidAU. Should it hold the same benchmarks of DavidAU´s?
The original source is here
https://huggingface.co/nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451
https://huggingface.co/nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-F32
maybe dont say The original source is here on DavidAU´s model page, will only amplify the confusion already created.
We should ask @DavidAU if he can upload f32 weights of the exact model in this repo (which is uncensored vs a bit uncensored with the one on nightmedia`s page, and maybe other differences I am not aware of.
I have added another source code repo here:
https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP
Thanks David, appreciate the shares of the enormous amount of training youve stacked up to get to this point.
AutoRound 4bit quant uploaded just now:
https://huggingface.co/DoktorMincs/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP-W4A16-AutoRound