license: apache-2.0
base_model:
- ConceptCountry/qwen-concept-27b
- huihui-ai/Huihui-Qwen3.8-27B-abliterated
tags: - gguf
- quantization
- unsloth-dynamic
- abliterated
- uncensored
Qwen Concept 27B — Abliterated (thinking-suppressed + refusal-removed)
Composite weight edit of
ConceptCountry/qwen-concept-27b
(Qwen3.8-27B, nine down-projections rewritten by weight-space programming: thinks
less, scores higher) on top of
huihui-ai/Huihui-Qwen3.8-27B-abliterated
(refusal direction removed, Arditi et al. 2024 style).
Result: a 27B hybrid-attention VLM that thinks less and refuses less — suited
to reverse-engineering and software-preservation research workloads.
- The concept edit is fully preserved: the 9 edited
mlp.down_projmatrices
(layers 32–48, per the upstreamEDIT_SPEC.json) are transplanted verbatim. - Everything else comes from the huihui abliterated base.
- Known tradeoff: those 9 matrices retain their original (un-ablated) weights, so
a small residual refusal surface exists through 9 of ~128 residual writers.
Empirically the model reasons freely through preservation-RE prompts.
Contents
| path | format | size |
|---|---|---|
bf16/ |
safetensors, bf16 | ~52 GB |
fp8/ |
FP8-E4M3 blockwise (128) | ~28 GB |
gguf/qwen-concept-27b-abliterated-BF16.gguf |
GGUF bf16 | ~51 GB |
gguf/qwen-concept-27b-abliterated-Q8_0.gguf |
GGUF Q8_0 | ~29 GB |
gguf/qwen-concept-27b-abliterated-UD-Q6_K_XL.gguf |
GGUF 7.40 bpw | ~25.3 GB |
gguf/qwen-concept-27b-abliterated-UD-Q5_K_XL.gguf |
GGUF 6.11 bpw | ~20.9 GB |
gguf/qwen-concept-27b-abliterated-UD-Q4_K_XL.gguf |
GGUF 5.14 bpw | ~17.6 GB |
gguf/qwen-concept-27b-abliterated-UD-Q3_K_XL.gguf |
GGUF 3.85 bpw | ~13.1 GB |
gguf/qwen-concept-27b-abliterated-UD-Q2_K_XL.gguf |
GGUF 2.87 bpw | ~9.8 GB |
gguf/mmproj-F16.gguf |
vision projector | ~0.9 GB |
UD rungs replay Unsloth Dynamic v3.0 per-tensor layer recipes
(unsloth/Qwen3.8-27B-GGUF), one hop from bf16, with an importance matrix
recomputed on these weights (2.8M-token chat-template calibration). MTP
weights retained for llama.cpp speculative decoding. No QAT/QAD — pure PTQ.
NVFP4 skipped (no Blackwell target).
Responsible use
Abliterated models have had safety guardrails surgically removed. Intended for
lawful research (software preservation, interoperability, security research) on
hardware you control. You are solely responsible for outputs.
License
Apache-2.0, inherited via Qwen3.8-27B and the upstream repos.