library_name: mlx
tags:
- mlx
- oq
- quantized
base_model: - DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP
DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF
My (Zviratko's) notes on this model
It scores marvelously on standard benchmarks
MMLU 88.0% 264 300 382.2 No
MMLU_PRO 70.0% 210 300 202.7 No
HUMANEVAL 89.6% 147 164 558.8 No
And it works very well with real tasks too.
It's not "more knowledgeable" than the base model or the datasets used to train the base, it's just more compatible with established frontier workflows and the way those model interact with the world in a harness.
Due to its finetunes and datasets used those used, it behaves much more like a frontier models in HOW this works - this is the effect of "distillation"
Some call it IP theft.
I would call it compatibility adaptation.
Since the labs that originaly created those models involed(Qwen/Alibaba, OpenAI, Anthropic and possibly Google Mind) used stolen, unlincensed and non-derivate data to train their models, then at most they can claim it breaks IP laws because their models already did it first.
They can't cleam IP rights to the model's outputs as those are uncopyrightable (at least by US law) No "laundering step" will help them, just as "money laundering" doesn't make drug money legal.
That's just the state of the landscape.
This note is NOT aimed at DavidAU and his colleagues who created this model, on the contrary, I believe it makes them immune from the frontier labs whose models provided them the data - as a paid service. You just can't copyright a generative AI output, nor should you try.
In reality, the correct answer to bigwigs out there alleging threats about distillation being theft is "Respectfully, your arguments are moronic, you a are demonstrably bad human being and maybe should stop pointing fingers." or just a simple "Fuck you, get back into your griefing hole and stay."
This is not legal advice, of course, use this at you own discretion. Use it well.
That's my 2 cents on the situation. I certainly don't feel morally bad for using those.
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-oQ6e-mtp
This model was quantized using oQ (oMLX v0.5.4.dev1) mixed-precision quantization.
Quantization details
- Model type: qwen3_5
- Bits: 6
- Group size: 64
- Format: MLX safetensors