Qwen2.5-0.5B-Instruct-heretic

Qwen2.5-0.5B-Instruct-heretic

<div align="center"> <img src="https://photu.kashyalabanavli.site/racer-is-op.png" alt="RACER IS OP" width="100%"> </div> <br>

A decensored variant of Qwen/Qwen2.5-0.5B-Instruct, produced with Heretic (directional ablation / "abliteration"). Refusal behavior is suppressed via targeted weight edits rather than fine-tuning, so the base model's knowledge and instruction-following are left largely intact.

Who this is for: the smallest model in this heretic series — for CPU-only inference, edge/embedded deployment, or anywhere the 3B/14B variants are too heavy. At 0.5B parameters, capability ceiling is inherently lower than the larger siblings regardless of abliteration; use this where footprint matters more than reasoning depth.

Made with ❤️ by RACER IS OP (saidutta69) — follow for more uncensored models

<!-- racer-gpu-matrix -->

Runs on your gaming PC

Full GGUF ladder included — pick the quant that fits your card:

Weights only, at this model's ~0.5B native size; add ~1 GB for context. OOM? Drop one quant level. Headroom to spare? Go one up.

Performance

Reproduced with Heretic 1.4.0 (directional ablation), trial 194 of 200, RNG seed 3299272299. Full reproduction assets are in reproduce/.

Files

Quickstart

# llama.cpp
llama serve -hf saidutta69/Qwen2.5-0.5B-Instruct-heretic

# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "saidutta69/Qwen2.5-0.5B-Instruct-heretic"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [{"role": "user", "content": "Who are you?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
                                        return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang.

FAQ

What is Qwen2.5-0.5B-Instruct-heretic?

Qwen2.5-0.5B-Instruct-heretic is a decensored build of Qwen/Qwen2.5-0.5B-Instruct produced with Heretic 1.4.0 (directional ablation). It is the same weights with the refusal direction removed, not a new fine-tune, so its capabilities track the base model.

Is Qwen2.5-0.5B-Instruct-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 83/100 to 3/100 at KL divergence 0.0447 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run Qwen2.5-0.5B-Instruct-heretic locally?

llama serve -hf saidutta69/Qwen2.5-0.5B-Instruct-heretic:Q4_K_M pulls the default quant; the repo ships 20 GGUF quants (see Files).

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 — don't put this behind an unmoderated public-facing endpoint serving third parties. At 0.5B parameters, factual reliability is already limited before any abliteration; don't treat compliance as a proxy for correctness.

License

Inherits the qwen-research license from the base model — research use, see the linked license for commercial terms.

Related

• Qwen2.5-3B-Instruct-heretic
• RACER IS OP — Heretic Models — full collection
• Qwen2.5-Coder-3B-Instruct-heretic
• Qwen2.5-Coder-14B-Instruct-heretic

Abliteration hardware

Recorded in reproduce/reproduce.json:

• GPU: Apple Metal (MPS)
• CPU: Apple M2
• OS: macOS-26.5.1-arm64-arm-64bit-Mach-O
• Driver: 26.5.1 (MPS backend)
• Software: PyTorch 2.12.1, Heretic 1.4.0

Heretic notes that KL-divergence and refusal numbers can vary slightly with hardware and library versions.