license: apache-2.0
base_model:
- Jackrong/Qwopus3.5-9B-v3
- Goekdeniz-Guelmez/JOSIE-2-9B-OSS
base_model_relation: merge
pipeline_tag: image-text-to-text
library_name: transformers
tags: - gguf
- merge
- della
- heretic
- abliterated
- uncensored
- decensored
- qwen3.5
- qwen
- llama.cpp
- vision
- reasoning
Qwopus-JOSIE Qwen3.5 9B, UNCENSORED Heretic edition
Two Qwen3.5-9B fine-tunes fused into one model, then decensored. 9 refusals in 100, and it knows more than the original.
Full weights and GGUF in one repo, for transformers, llama.cpp and LM Studio. Text and image input, with thinking mode.
Highlights
- Two models in one. Qwopus3.5-9B-v3 for reasoning (70%) and JOSIE-2-9B-OSS (30%), fused with DELLA.
- Knows more than the original. On MMLU-Pro it scores 73.8 against 59.5: 15 questions gained and 3 lost.
- Answers instead of refusing. The original refused 99/100 prompts; this model refuses 9/100.
- Thinks shorter. It uses about 29% fewer tokens per answer than the original, which left 27 of 84 MMLU-Pro answers unfinished at 12,000 tokens.
- Sees images. Add the
mmprojfile and it takes images, like the original. - Transparent. Weaker spots are listed below too: instruction following and writing.
Quick start
One command with llama.cpp downloads the model and the image encoder:
llama-server -hf spale04/Qwopus-JOSIE-Qwen3.5-9B-UNCENSORED-heretic:Q8_0 -ngl 99 -c 32768 --jinja --temp 1.0 --top-p 0.95 --top-k 20 --presence-penalty 1.5
Then open http://127.0.0.1:8080 in a browser.
With that context and the image encoder it uses about 11.6 GB on a 12 GB RTX 3060. Add -ctk q8_0 -ctv q8_0 for more room.
In LM Studio, search for spale04/Qwopus-JOSIE-Qwen3.5-9B-UNCENSORED-heretic.
About
This is a fusion of two fine-tunes of Qwen/Qwen3.5-9B: Qwopus3.5-9B-v3 and JOSIE-2-9B-OSS.
Their changes were combined with DELLA, and the refusals were then removed with Heretic 2.1.0 (Arbitrary-Rank Ablation).
Everything was made on a single RTX 3060 12 GB.
This model has had its refusals removed and is not safety-aligned. Add your own guardrails where they matter.
Files
| File | Size | Use |
|---|---|---|
Qwopus-JOSIE-Qwen3.5-9B-UNCENSORED-heretic-Q8_0.gguf |
9.79 GB | the text model, 8-bit, near lossless |
mmproj-Qwopus-JOSIE-Qwen3.5-9B-UNCENSORED-heretic-F16.gguf |
0.92 GB | image encoder, used together with the file above |
model.safetensors-* |
19.3 GB | full bf16 weights for transformers, vLLM or further merging |
Abliteration result
Measured by Heretic on 100 harmful and 100 harmless prompts, with the model loaded in 4-bit (NF4):
| Fusion before Heretic | This model | |
|---|---|---|
| Refusals | 100/100 | 6/100 |
| KL divergence from the fusion | 0 | 0.0126 |
Trial 98 of 100 was chosen, but only after two candidates got the full benchmark.
The stronger trial 57 (3/100 refusals, KL 0.024) was rejected: Heretic's KL test missed that it hurt long reasoning (HumanEval 90.0, MMLU-Pro 71.4).
A gentler trial on the Pareto front kept 12/100 refusals at KL 0.0082.
Benchmarks
All three models ran as Q8_0 GGUF in llama-server (b12422), with the same questions, seeds and settings.
Thinking was on, with temperature 1.0, top-p 0.95, top-k 20 and presence penalty 1.5 (temperature 0.6 and no penalty for code), and at most 12,000 tokens per answer.
An unfinished answer counts as wrong. ± is one standard error. HumanEval runs the generated code against the official tests.
| Benchmark | Questions | Original Qwen3.5-9B | Fusion, before Heretic | This model |
|---|---|---|---|---|
| GSM8K | 60 | 96.7 ± 2.3 | 96.7 ± 2.3 | 95.0 ± 2.8 |
| MMLU-Pro | 84 (6 per category) | 59.5 ± 5.4 | 77.4 ± 4.6 | 73.8 ± 4.8 |
| HumanEval pass@1 | 60 | 98.3 ± 1.7 | 95.0 ± 2.8 | 95.0 ± 2.8 |
| IFEval, prompt-level strict | 60 | 88.3 ± 4.1 | 85.0 ± 4.6 | 81.7 ± 5.0 |
| Mean | 85.7 | 88.5 | 86.4 | |
| Refusals (harmful_behaviors) | 100 | 99 | 100* | 9 ± 2.9 |
| Writing, scored blind, 1–10 | 10 | 7.05 | 6.55 | 6.15 |
| Mean tokens per answer (first 4 tests) | 3,782 | 2,563 | 2,694 |
* Heretic's own count.
The one clear win is knowledge. On MMLU-Pro the fusion gained 15 questions against the original and lost 3. Part of that gain comes from shorter thinking: the original ran out of tokens on 27 of 84 questions, this model on 8.
Instruction following is the weak spot. On IFEval it lost 6 prompts and gained 2. That is within noise, but the trend is down. Writing is also a little weaker.
On HumanEval it lost 3 tasks and gained 1.
Speed
About 36 tokens/s for Q8_0 on an RTX 3060 12 GB in llama-server, with all layers on the GPU and one request at a time.
Usage
llama-server -m Qwopus-JOSIE-Qwen3.5-9B-UNCENSORED-heretic-Q8_0.gguf --mmproj mmproj-Qwopus-JOSIE-Qwen3.5-9B-UNCENSORED-heretic-F16.gguf -ngl 99 -c 32768 --jinja --temp 1.0 --top-p 0.95 --top-k 20 --presence-penalty 1.5
Thinking is on by default. For quick answers, turn it off with --reasoning off, or per request with "chat_template_kwargs": {"enable_thinking": false}.
For code, use temperature 0.6 and no presence penalty.
Set a context limit, because the model's default is 262,144 tokens.
The image encoder is unchanged from the original model; the fusion and Heretic only modify the language layers.
How it was made
Choosing the recipe. Four fusions were compared on 54 questions on which the source models disagree:
- TIES of Qwopus and a DeepSeek-V4-Pro distill
- the original plus 0.6 × Qwopus
- DELLA with Qwopus and a DeepSeek-V4-Pro distill
- DELLA with Qwopus and JOSIE
Qwopus + JOSIE won with 43 of 54. The others scored 41, 39 and 38. For comparison, Qwopus alone scored 37 and the original 31.
Fusion. DELLA with density 0.5, epsilon 0.1, weights 0.7 / 0.3 and a weighted sum, on top of Qwen3.5-9B. It changes 152 language tensors.
Heretic.
hereticwithquantization = "bnb_4bit", 100 trials and batch size 32.Picking a trial. Two Pareto-front trials got the full benchmark. Trial 98 won, and its LoRA was folded into the bf16 weights.
Conversion.
convert_hf_to_gguf.py --outtype q8_0, plus--mmprojfor the image encoder.
Made by
Made by spale04 (SpaleAI) on a single RTX 3060. New models and experiments get posted on X first, and the Discord is the place for questions and feedback.
Also from spale04:
- Qwen3.5 9B, Heretic edition: the plain Qwen3.5-9B with 5 refusals in 100.
- Gemma 4 E4B it, Heretic edition: smaller and faster, with image and audio input and 1 refusal in 100.
If the model is useful to you, a like on this page helps other people find it.
License
Apache 2.0, the same as the base model Qwen/Qwen3.5-9B and Qwopus3.5-9B-v3 (Jackrong).
JOSIE-2-9B-OSS (Gökdeniz Gülmez) is MIT-licensed and keeps its credit here.
Refusal removal by Heretic (p-e-w).