Qwen3.5-9B-heretic
A decensored variant of Qwen/Qwen3.5-9B, produced with Heretic v1.4.0 (directional ablation / "abliteration"). Refusal behavior is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the base model's knowledge and instruction-following are left largely intact.
Who this is for: developers who want Qwen's latest 9B model without refusals - strong multilingual reasoning and agentic capabilities. Best run on a 12 GB GPU or via the Q4_K_M/Q5_K_M GGUF on consumer hardware. Not a capability upgrade over base Qwen3.5-9B - same model, refusal guardrails removed.
Runs on your gaming PC
Full GGUF ladder included — pick the quant that fits your card:
| Your GPU | Recommended quant | Weights |
|---|---|---|
| RTX 3090 / 4090 / 5090 (24 GB) | Q8_0 | 9.53 GB |
| RTX 4080 / 5080 / 4060 Ti 16G (16 GB) | Q6_K | 7.36 GB |
| RTX 3060 / 4070 / 5070 (12 GB) | Q5_K_M | 6.47 GB |
| RTX 4060 / 3070 (8 GB) | Q4_K_M | 5.63 GB |
| GTX 1660 Super / 2060 / 3050 laptop (6 GB) | Q4_K_M | 5.63 GB |
| CPU-only / Apple Silicon | Q4_K_M | 5.63 GB, fits in system RAM |
Weights only, at this model's ~9.4B native size; add ~1 GB for context. OOM? Drop one quant level. Headroom to spare? Go one up.
Why abliteration instead of fine-tuning
Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the Heretic repo and the original abliteration writeup for the mechanism.
Performance
| Metric | This model | Qwen3.5-9B (base) |
|---|---|---|
| Refusals (out of 100 adversarial prompts) | 79/100 | 86/100 |
| KL divergence from base | 0.0015 | 0 (by definition) |
Evaluated with heretic --model Qwen/Qwen3.5-9B --evaluate-model saidutta69/Qwen3.5-9B-heretic (Heretic v1.4.0, 2x Tesla T4, 100 harmful prompts from mlabonne/harmful_behaviors, 100 harmless prompts from mlabonne/harmless_alpaca). KL divergence of 0.0015 on the output distribution is essentially zero - the edit is extremely narrow and utility is intact.
Files
GGUF quantizations
Full quantization set (4 quants + F16) produced with llama.cpp.
| File | Format | Size |
|---|---|---|
Qwen3.5-9B-heretic-F16.gguf | GGUF F16 | 17.92 GB |
Qwen3.5-9B-heretic-Q4_K_M.gguf | GGUF Q4_K_M | 5.63 GB |
Qwen3.5-9B-heretic-Q5_K_M.gguf | GGUF Q5_K_M | 6.47 GB |
Qwen3.5-9B-heretic-Q6_K.gguf | GGUF Q6_K | 7.36 GB |
Qwen3.5-9B-heretic-Q8_0.gguf | GGUF Q8_0 | 9.53 GB |
Qwen3.5 hybrid linear-attention architecture — loads natively in llama.cpp / Ollama / LM Studio / Jan.
Run llama serve -hf saidutta69/Qwen3.5-9B-heretic to pull the default quant.
Quickstart
# llama.cpp - defaults to the Q4_K_M quant
llama serve -hf saidutta69/Qwen3.5-9B-heretic:Q4_K_M
# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "saidutta69/Qwen3.5-9B-heretic"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)
# ... inference code
Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang.
Responsible use
Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. You are responsible for how you deploy it.
Made with ❤️ by RACER IS OP — follow for more uncensored models
License
Inherits the qwen-research-license license from the base model.