Qwen3.5-2B-heretic

A decensored variant of Qwen/Qwen3.5-2B, 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 fast Qwen3.5 model without refusals - the 2B 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 Qwen3.5-2B - same model, refusal guardrails removed.

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 |

|---|---|---|

| Qwen3.5-2B-heretic-F16.gguf | GGUF F16 | 3.52 GB |

| Qwen3.5-2B-heretic-Q2_K.gguf | GGUF Q2_K | 924 MB |

| Qwen3.5-2B-heretic-IQ3_S.gguf | GGUF IQ3_S | 1002 MB |

| Qwen3.5-2B-heretic-Q3_K_S.gguf | GGUF Q3_K_S | 973 MB |

| Qwen3.5-2B-heretic-Q3_K_M.gguf | GGUF Q3_K_M | 1.02 GB |

| Qwen3.5-2B-heretic-Q3_K_L.gguf | GGUF Q3_K_L | 1.08 GB |

| Qwen3.5-2B-heretic-IQ4_XS.gguf | GGUF IQ4_XS | 1.12 GB |

| Qwen3.5-2B-heretic-Q4_K_S.gguf | GGUF Q4_K_S | 1.13 GB |

| Qwen3.5-2B-heretic-Q4_0.gguf | GGUF Q4_0 | 1.12 GB |

| Qwen3.5-2B-heretic-Q4_1.gguf | GGUF Q4_1 | 1.20 GB |

| Qwen3.5-2B-heretic-Q4_K_M.gguf | GGUF Q4_K_M | 1.19 GB |

| Qwen3.5-2B-heretic-Q5_K_S.gguf | GGUF Q5_K_S | 1.28 GB |

| Qwen3.5-2B-heretic-Q5_K_M.gguf | GGUF Q5_K_M | 1.31 GB |

| Qwen3.5-2B-heretic-Q6_K.gguf | GGUF Q6_K | 1.45 GB |

| Qwen3.5-2B-heretic-Q8_0.gguf | GGUF Q8_0 | 1.87 GB |

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

Run llama serve -hf saidutta69/Qwen3.5-2B-heretic to pull the default quant.

Quickstart

# llama.cpp - defaults to the Q4_K_M quant
llama serve -hf saidutta69/Qwen3.5-2B-heretic:Q4_K_M

# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "saidutta69/Qwen3.5-2B-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.