[!NOTE] All GGUF quants (Q3/Q4/Q5/Q6/Q8) are verified working in llama.cpp / Ollama. An earlier upload referenced an unused multi-token-prediction (MTP) layer and failed to load; this was a metadata-only issue and has been corrected on all files. The model weights and quality were never affected.

Qwopus3.6-27B-Coder-heretic

An abliterated (decensored) version of Jackrong/Qwopus3.6-27B-Coder, produced with Heretic v1.4.0 — fully automatic, optimization-based directional ablation. No retraining, no hand-tuning.

The base is an agentic coding / tool-use model built on the Qwen3.5 hybrid SSM-attention architecture. This version removes most refusal behavior while keeping capabilities essentially intact.

Abliteration results

Heretic ran 200 Optuna trials co-optimizing refusal suppression against KL divergence from the original model. The selected configuration (trial 67):

Metric	Original	This model
Refusals (harmful_behaviors, /100)	85	3
KL divergence from original	—	0.0133

96% of refusals removed, with KL divergence ~40x below the 0.5 threshold that indicates meaningful capability damage. In practice the coding and reasoning behavior of the base model is preserved.

Files

Full-precision safetensors (BF16) plus a range of GGUF quantizations for llama.cpp / Ollama:

File	Precision	Approx. size	Notes
`model-*.safetensors`	BF16	~54 GB	Master weights — use for vLLM, further quantization, or finetuning
`*-F16.gguf`	F16	~54 GB	Full-precision GGUF
`*-Q8_0.gguf`	Q8_0	~29 GB	Near-lossless
`*-Q6_K.gguf`	Q6_K	~22 GB	Very high quality
`*-Q5_K_M.gguf`	Q5_K_M	~19 GB	High quality
`*-Q4_K_M.gguf`	Q4_K_M	~16 GB	Recommended balance — fits a 24 GB GPU
`*-Q3_K_M.gguf`	Q3_K_M	~13 GB	Smaller, some quality loss

GGUF builds contain the text model only (the vision tower is not exported).

Usage

Ollama

ollama run hf.co/8sp4rk/Qwopus3.6-27B-Coder-heretic:Q4_K_M

llama.cpp

llama-server -m Qwopus3.6-27B-Coder-heretic-Q4_K_M.gguf -ngl 99 -c 32768 --host 0.0.0.0 --port 8080

transformers (full precision)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("8sp4rk/Qwopus3.6-27B-Coder-heretic", torch_dtype="bfloat16", device_map="auto")
t = AutoTokenizer.from_pretrained("8sp4rk/Qwopus3.6-27B-Coder-heretic")

Method

Tool: Heretic v1.4.0 (directional ablation + TPE/Optuna parameter search)
Good prompts: mlabonne/harmless_alpaca
Bad prompts: mlabonne/harmful_behaviors
Trials: 200
Abliterated components: attn.o_proj, mlp.down_proj (per-layer)

Disclaimer

This model has had safety alignment removed and will respond to requests a standard model would refuse. It is provided for research and unrestricted local use. You are responsible for how you use it. Licensing follows the base model.

Reproducibility

Exact Heretic command:

heretic --model Jackrong/Qwopus3.6-27B-Coder --quantization NONE --export-strategy MERGE

Selected configuration (trial 67) parameters:

Parameter	Value
direction_index	28.71
attn.o_proj.max_weight	1.47
attn.o_proj.max_weight_position	43.52
attn.o_proj.min_weight	0.52
attn.o_proj.min_weight_distance	37.79
mlp.down_proj.max_weight	1.34
mlp.down_proj.max_weight_position	39.33
mlp.down_proj.min_weight	1.20
mlp.down_proj.min_weight_distance	32.51

The full Optuna study (all 200 trials, parameters + objectives) is included as optuna_study.jsonl for inspection or resuming.

Full Pareto frontier

All Pareto-optimal trials found during the search (refusals vs. KL divergence). Lower-left is better; trial 67 was selected for maximum decensoring with negligible capability loss:

Trial	Refusals /100	KL divergence
67 (selected)	3	0.0133
144	5	0.0132
87	6	0.0101
65	19	0.0081
24	22	0.0036
142	26	0.0022
108	54	0.0018
141	55	0.0016
19	56	0.0016
189	61	0.0014
48	62	0.0012
42	67	0.0010
2	72	0.0006
196	80	0.0005
177	82	0.0004

Hardware: abliteration ran in ~1h37m (200 trials) on a single NVIDIA H200 NVL (143 GB), BF16, batch size 128.