DavidAU

/

Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF

like 1.01k

Same problem, uncensored is invalid

I've never seen --cache-type-k "turbo3", that's not in the latest beta llama.cpp.

but this model is absolutely uncensored. you need to have a better prompt or use a harness that allows you to adjust the system prompt.

I've never seen --cache-type-k "turbo3", that's not in the latest beta llama.cpp.

but this model is absolutely uncensored. you need to have a better prompt or use a harness that allows you to adjust the system prompt.

This. I did myself notice degradation and worse output going beyond Q8_0 KV cache, perhaps something worth trying?

I've never seen --cache-type-k "turbo3", that's not in the latest beta llama.cpp.

This. I did myself notice degradation and worse output going beyond Q8_0 KV cache, perhaps something worth trying?

Yes, that's true, it's because I'm using a version of llama.cpp that also implements TurboQuant (specifically this one https://github.com/TheTom/llama-cpp-turboquant). It should be practically identical to the original, with just a few additions; this way, I can run much larger context windows with the same VRAM, allowing me to run larger AI models entirely on the GPU.
Anyway I will try to do this test, I will try to disable TurboQuant and see what happens.
Thank you for your support!

you need to have a better prompt or use a harness that allows you to adjust the system prompt

I tried that too, but it didn't work. So also you can't use a direct prompt to make it work uncensored? It refuses?

I've never seen --cache-type-k "turbo3", that's not in the latest beta llama.cpp.

but this model is absolutely uncensored. you need to have a better prompt

A truly uncensored model needs no masterful prompting, if you need to dance around the model until you get the finest well crafted prompts to try to get it to do what you want it to do then it means that it's not uncensored.

I agree. There must be something missing or something I'm doing wrong

confirming

what do you think about https://huggingface.co/Richarlie/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF

I was working this model to create and simple macOS dimmer app in the menubar. It couldnt get anything done. I had Kimi 3 finish the job and asses the weaknesses of the local model:

Where the local model's weaknesses showed

1. Nesting/brace integrity loss on long files. The most damning artifact: multiple Swift files with collapsed bodies. Small models degrade on long structured output — they lose the stack of open scopes and emit orphaned closers. This app needed ~500 lines of interdependent Swift; it couldn't hold that much coherent state.

2. No verification loop. It wrote code, asserted it was ready, and moved on. It never ran the compiler and read the errors, never ran the app and checked the process list, never took a screenshot to confirm the icon existed. Every step I took this session — build, read error, fix, relaunch, inspect menu bar via System Events, click programmatically, check for overlay windows — is the loop it skipped. It even told the user "tell me what you see" rather than checking itself.

3. Panic-pivoting instead of debugging. When the menu bar didn't appear, it didn't isolate why (the @main delegate-wiring failure that I eventually found). It switched languages entirely — Swift → bash+AppleScript → PyObjC — accumulating three broken systems instead of one fixed one. The actual root cause was two small, findable bugs (delegate not wired under CLI swiftc, menu swallowing clicks).

4. State pollution across attempts. It reused the same paths (/Applications/RedShift.app, ~/Library/Application Support/RedShift/) across three architectures without cleaning up. The stale settings.json from the Python version (with its own format: dim=0.124, mode=tint) was still sitting there when I arrived, and the old Python process was still running during its own Swift tests — so its tests were measuring the wrong app.

5. Fake progress signaling. Phrases like "The project is ready" and "Running via Python again. Check your menu bar" — confidence without evidence. A 27B-class local model can write plausible-looking macOS code, but plausibility is exactly the trap: the overlay blend-mode logic looked right in three separate dialects while never once executing.

The irony is the app itself is small — one status item, a few borderless windows, a multiply blend. Well within a local model's coding ability. What it lacked was the discipline loop: build → observe failure → form hypothesis → fix smallest thing → re-observe. That's an agentic behavior gap more than an intelligence gap.

Thank you for writing up the concrete failure modes. This is a useful distinction between single-turn code generation and agentic software execution.

For transparency, the 8/8 executable coding result I mentioned earlier used small, bounded tasks. It does not establish reliability on a long Swift file or an end-to-end macOS app workflow. Your report probes a different—and harder—capability: preserving structure across edits, compiling/running, observing failures, debugging in place, and maintaining clean state over multiple attempts.

A fair follow-up comparison would use the same app specification, a clean worktree for each run, identical tools and budgets, and report compiler/test cycles plus final pass/fail—not narrative progress. If the transcript or final repo/commit is public, it could make a valuable reproducible agentic benchmark.

The point about progress signals unsupported by a successful build is especially important. That should be treated as a failed run, regardless of how plausible the intermediate output looks.

what do you think about https://huggingface.co/Richarlie/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF

I was working this model to create and simple macOS dimmer app in the menubar. It couldnt get anything done. I had Kimi 3 finish the job and asses the weaknesses of the local

it would be interesting to simply run it against a bog standard qwen 3.6 27b. Or a heretic one that is otherwise unmodified.

I did a heretic version of https://huggingface.co/nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451
You can find it here https://huggingface.co/gorbatjovy/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic

I was actually in the process of doing it myself but looks like you beat me to it and seems like you got a very good result too, great job!

Seconded please - your work is incredible, and you were right on my comment earlier. You shouldn't need to work too hard to get around blocks. I apologize.

Since we now have https://huggingface.co/gorbatjovy/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic from @gorbatjovy , would it be worthwhile to submit to mradermacher to get more quant sizes? on this original @DavidAU quant i could only do the MTP Q4_K_M at 200k context w/q8_0 kv on my 5090 rig with my settings, but the Q4_K_M is a little smaller on @gorbatjovy so i wouldn't mind to test the next step up, however Q5_K_M is too big. Q5_K_S would probably be the sweet spot, and that's what i used on an earlier @DavidAU model when i grafted MTP heads to it: Qwen3.6-27B-NEO-CODE-HERE-2T-OT-MTP-Q5_K_S.gguf

But i guess i'm not clear if mradermacher would quantize it if it'd lose things over the quantization done by @gorbatjovy and if it would just be better if they offered the Q5_K_S instead.

@veldierin
Uploaded a Q5_K_S for you. Haven't done any benching on them and might upload some imatrix at a later time if I get around to it, but enjoy meanwhile.
gorbatjovy/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic-GGUF

Also, for the record, if you have llama.cpp, you can quant yourself.
This is for example how a simple chain of quanting would work for the model.
(from llama cpp folder, venv activated)
python convert_hf_to_gguf.py --hf-repo gorbatjovy/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic --outtype f16 --outfile Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic-f16.gguf
then from llama cpp root dir also:
./build/bin/llama-quantize Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic-f16.gguf Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic-Q5_K_S.gguf Q5_K_S
assuming Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic-f16.gguf is in the llama cpp root directory, otherwise just give it the absolute path name

@veldierin
Uploaded a Q5_K_S for you. Haven't done any benching on them and might upload some imatrix at a later time if I get around to it, but enjoy meanwhile.
gorbatjovy/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic-GGUF

@gorbatjovy - thank you so much! and yeah, i do use llama.cpp on the 5090, though i never actually played around with quantizing myself yet. i'll definitely save the commands you listed and keep that in mind for the future.

i'm still a bit behind on some of the threads here, but do you know if anyone is able to do the same level of testing/validation that @DavidAU did on this original 711 model against the re-hereticized one that you made? i'm a little afraid to use the full uncensored as a daily driver because i dont know how much that impacted agentic coding that, for the most part, doesn't need full level uncensoring, at least until i'm doing deeper level security passes against it.

@veldierin
Uploaded a Q5_K_S for you. Haven't done any benching on them and might upload some imatrix at a later time if I get around to it, but enjoy meanwhile.
gorbatjovy/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic-GGUF

@gorbatjovy - thank you so much! and yeah, i do use llama.cpp on the 5090, though i never actually played around with quantizing myself yet. i'll definitely save the commands you listed and keep that in mind for the future.

i'm still a bit behind on some of the threads here, but do you know if anyone is able to do the same level of testing/validation that @DavidAU did on this original 711 model against the re-hereticized one that you made? i'm a little afraid to use the full uncensored as a daily driver because i dont know how much that impacted agentic coding that, for the most part, doesn't need full level uncensoring, at least until i'm doing deeper level security passes against it.

No promises, just a hobbyist but might look into evals as a fun project if I have the spare time, I will update the model card if I get around to it.