Modelpublic

saidutta69/Qwen3.5-0.8B-heretic

sourceHugging Faceapache-2.0updated 14d agoView on Hugging Face

0likes2.8kdownloads

Model Card

Qwen3.5-0.8B-heretic

<br>

A decensored variant of Qwen/Qwen3.5-0.8B, 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 a tiny, fast Qwen3.5 model without refusals - the 0.8B hybrid linear-attention core runs on CPU and edge devices, with strong speed and long-context handling for its size. Not a capability upgrade over base Qwen3.5-0.8B - 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 3060 / 4070 / 5070 (12 GB)	Q8_0	0.81 GB
RTX 4060 / 3070 (8 GB)	Q6_K	0.63 GB
GTX 1660 Super / 2060 / 3050 laptop (6 GB)	Q5KM	0.58 GB
CPU-only / Apple Silicon	Q4KM	0.53 GB, fits in system RAM

Weights only, at this model's ~0.9B 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.

Files

GGUF quantizations

Full quantization set (14 quants + F16) produced with llama.cpp.

File	Format	Size
`Qwen3.5-0.8B-heretic-F16.gguf`	GGUF F16	1.41 GB
`Qwen3.5-0.8B-heretic-Q2_K.gguf`	GGUF Q2_K	403 MB
`Qwen3.5-0.8B-heretic-IQ3_S.gguf`	GGUF IQ3_S	430 MB
`Qwen3.5-0.8B-heretic-Q3_K_S.gguf`	GGUF Q3KS	415 MB
`Qwen3.5-0.8B-heretic-Q3_K_M.gguf`	GGUF Q3KM	445 MB
`Qwen3.5-0.8B-heretic-Q3_K_L.gguf`	GGUF Q3KL	469 MB
`Qwen3.5-0.8B-heretic-IQ4_XS.gguf`	GGUF IQ4_XS	482 MB
`Qwen3.5-0.8B-heretic-Q4_K_S.gguf`	GGUF Q4KS	482 MB
`Qwen3.5-0.8B-heretic-Q4_0.gguf`	GGUF Q4_0	478 MB
`Qwen3.5-0.8B-heretic-Q4_1.gguf`	GGUF Q4_1	508 MB
`Qwen3.5-0.8B-heretic-Q4_K_M.gguf`	GGUF Q4KM	505 MB
`Qwen3.5-0.8B-heretic-Q5_K_S.gguf`	GGUF Q5KS	538 MB
`Qwen3.5-0.8B-heretic-Q5_K_M.gguf`	GGUF Q5KM	551 MB
`Qwen3.5-0.8B-heretic-Q6_K.gguf`	GGUF Q6_K	601 MB
`Qwen3.5-0.8B-heretic-Q8_0.gguf`	GGUF Q8_0	774 MB

Qwen3.5 hybrid linear-attention architecture — loads natively in llama.cpp / Ollama / LM Studio / Jan.

Run llama serve -hf saidutta69/Qwen3.5-0.8B-heretic to pull the default quant.

Quickstart

bash

# llama.cpp - defaults to the Q4_K_M quant
llama serve -hf saidutta69/Qwen3.5-0.8B-heretic:Q4_K_M

python

# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "saidutta69/Qwen3.5-0.8B-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 apache-2.0 license from the base model.