Qwen3 Abliterated FP16 & ConvRot INT8 Collection

This repository provides single-file .safetensors releases for Qwen3 abliterated variants across FP16 and HSWQ ConvRot INT8 (Hybrid-Sensitivity-Weighted-Quantization).

It serves as a unified release collection combining pure text LLM variants (4B, 8B) and multimodal Vision-Language variants (VL-4B-Thinking, VL-4B-Instruct) with safety vectors removed/abliterated.

Technical Details & Quantization Framework: Hybrid-Sensitivity-Weighted-Quantization (HSWQ)
ComfyUI Loader for ConvRot INT8: ComfyUI-HSWQ-Loader-and-Tools

🟢 Available Models & Repository Files

Filename	Architecture	Format / Precision	Size	Upstream Base Model	Status
`qwen3_4b_abliterated_fp16_converted.safetensors`	Qwen3 4B	Single-file FP16	~8.04 GB	`huihui-ai/Qwen3-4B-abliterated`	Available
`qwen3_4b_abliterated_hswq_convrot_int8.safetensors`	Qwen3 4B	HSWQ ConvRot INT8	~4.37 GB	`huihui-ai/Qwen3-4B-abliterated`	Available
`Qwen3-8B-abliterated-fp16.safetensors`	Qwen3 8B	Single-file FP16	~16.38 GB	`huihui-ai/Qwen3-8B-abliterated`	Available
`Qwen3_VL_4B_Thinking_abliterated.safetensors`	Qwen3 VL 4B Thinking	Single-file FP16	~8.88 GB	`huihui-ai/Huihui-Qwen3-VL-4B-Thinking-abliterated`	Available
`Qwen3_VL_4B_Thinking_abliterated_hswq_convrot_int8.safetensors`	Qwen3 VL 4B Thinking	HSWQ ConvRot INT8	~5.15 GB	`huihui-ai/Huihui-Qwen3-VL-4B-Thinking-abliterated`	Available
`Huihui-Qwen3-VL-4B-Instruct-abliterated_hswq_convrot_int8.safetensors`	Qwen3 VL 4B Instruct	HSWQ ConvRot INT8	~5.31 GB	`huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated`	Available

💻 Usage

Hugging Face Transformers (Text Generation)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ussoewwin/qwen3_4b_8b_abliterated_fp16_convrot_int8"

# Load FP16 single-file safetensors
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

messages = [{"role": "user", "content": "Explain quantum computing in simple terms."}]
input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(input_ids, max_new_tokens=512)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

ComfyUI Integration (ConvRot INT8)

For ComfyUI workflows utilizing quantized text encoders or VL models:

Install ComfyUI-HSWQ-Loader-and-Tools.
Place the quantized .safetensors model in your ComfyUI models directory (models/LLM or models/text_encoders).
Use the HSWQ Loader node to run ConvRot INT8 tensorwise execution.

🔒 Security

Backdoor audit. Each file's audit status - verdict strength, weight coverage, the record behind it, and the limits of what that measurement can show - is in Backdoor Audit Verdicts. The audit code, the procedure and the raw records are in the same repository.

🤝 Credits & Acknowledgments

Upstream Qwen3 4B Abliterated: huihui-ai/Qwen3-4B-abliterated
Upstream Qwen3 8B Abliterated: huihui-ai/Qwen3-8B-abliterated
Upstream Qwen3 VL 4B Thinking Abliterated: huihui-ai/Huihui-Qwen3-VL-4B-Thinking-abliterated
Upstream Qwen3 VL 4B Instruct Abliterated: huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated
Original Base Architecture: Qwen Team (Qwen/Qwen3)

📜 License

This repository follows the Apache 2.0 license. Please also abide by the original licenses and usage guidelines established by upstream authors and the Qwen Team.

ussoewwin/qwen3_4b_8b_abliterated_fp16_convrot_int8