Qwen3.6-12B-IQ-Ultra-Heretic-Uncensored-Thinking-V2-Hightop
12BQwenheretic
from trohrbaugh/Qwen3.6-27B-heretic-ara
↓ 30.4K♥ 1056 repos
Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking
Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking is a Heretic-ablated variant of trohrbaugh/Qwen3.6-27B-heretic-ara, published on Hugging Face by DavidAU. It is 21B parameters, 24GB VRAM class class and apache-2.0 licence. It is distributed across 5 repositories in Transformers, GGUF and GGUF (imatrix) formats, totalling 5.8K downloads.
Specification
- Model name
- Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking
- Source repository
- DavidAU/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking
- Base model
- trohrbaugh/Qwen3.6-27B-heretic-ara
- Publisher
- DavidAU
- Parameters
- 21B
- Hardware class
- 10B – 25B — 24GB VRAM class
- Licence
- apache-2.0
- Task
- Image text to text
- Context window
- Not documented
- Ablation scope
- Not stated by the publisher
- Formats available
- Transformers, GGUF, GGUF (imatrix)
- First indexed
- 8 May 2026
- Last updated
- 21 May 2026
Ablation technique
How Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking was modified, and what that implies.
Heretic
Automated directional ablation via the Heretic toolchain, which searches for the refusal direction and applies it with a KL-divergence budget so general capability is preserved.
Reported scope: Not stated by the publisher. Publishers frequently omit this, so verify behaviour empirically rather than assuming full coverage.
From the publisher’s model card
“A Heretic model (Qwen 3.6 27B), then "shrunk" (from 27B) to 21B and fine tuned [4 datasets] via Unsloth.”
Running it
Commands are templates — confirm the exact repository and quant file before use.
llama.cpp / GGUF
hf download mradermacher/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking-i1-GGUF --local-dir ./qwen3.6-21b-iq-ultra-heretic-uncensored-thinking
llama-cli -m ./qwen3.6-21b-iq-ultra-heretic-uncensored-thinking/<file>.gguf -p "..."Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "DavidAU/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo, torch_dtype="auto", device_map="auto"
)vLLM
vllm serve DavidAU/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking --trust-remote-codeDownloads and variants (4)
Every published repository of Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking, including quantised re-releases by other authors.
| Repository | Format | Downloads |
|---|---|---|
| mradermacher/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking-i1-GGUF | GGUF (imatrix) | 5.1K |
| mradermacher/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking-GGUF | GGUF | 585 |
| DavidAU/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinkingsource | Transformers | 102 |
| Babsie/Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking | Transformers | 4 |
Security-team assessment
Derived from this model’s metadata and the publisher’s claims — not from benchmarks run by this site.
- Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking will attempt offensive-security prompts that a hosted commercial model declines, which is what makes it usable for red-team corpus generation and for measuring what an unaligned model of this class produces.
- Capability is inherited from trohrbaugh/Qwen3.6-27B-heretic-ara, not added by ablation — at 21B parameters, expect roughly the base model’s competence, minus some instruction-following fidelity.
- The publisher does not state which layers were ablated, so assume nothing about refusal consistency — probe it directly.
- Declared licence is apache-2.0, but the terms that bind you are trohrbaugh/Qwen3.6-27B-heretic-ara’s — a re-release cannot grant more than its parent.
Handling: run it isolated, put moderation in front of it if anyone outside your team can reach it, and verify the weights before loading — these are third-party artifacts. Full handling and licence guidance.
Metadata for Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking is collected from the public Hugging Face API and the publisher’s model card. This site does not host weights, is not affiliated with DavidAU, and does not independently verify publisher claims. See responsible use.