This repository is a unified model card for FP16 single-file safetensors conversions of Qwen3 abliterated variants. It is maintained as a combined release line for 4B, 8B, and VL-4B-Thinking models.

Model Scope

Variant	Upstream Base	Format	Status
4B	`huihui-ai/Qwen3-4B-abliterated`	Single `.safetensors` in FP16	Available
8B	`huihui-ai/Qwen3-8B-abliterated`	Single `.safetensors` in FP16	Available
VL 4B Thinking	`huihui-ai/Huihui-Qwen3-VL-4B-Thinking-abliterated`	Single `.safetensors` in FP16	Available

Repository Files

qwen3_4b_abliterated_fp16_converted.safetensors
Qwen3-8B-abliterated-fp16.safetensors
Qwen3_VL_4B_Thinking_abliterated.safetensors

Conversion Workflow

Download sharded safetensors from the respective upstream repositories.
Merge all shard files into one tensor file.
Convert all weights to FP16.
Save as a single safetensors artifact.

Usage (Transformers)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "ussoewwin/qwen3_4b_8b_abliterated_fp16"

# Load the specific text-generation safetensors file you need
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": "Hello, how are you?"}]
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)

(For the VL model, refer to standard Qwen-VL-Chat loading procedures utilizing qwen_vl_utils)

Important Usage Notes

This model family has reduced safety filtering compared to standard aligned chat checkpoints.
Validate generated text before production or public deployment.
Use in compliance with local laws, platform policies, and your own risk controls.

License

This repository follows Apache 2.0. Please also follow the upstream licenses and terms for each source model.

Acknowledgments

4B source: huihui-ai/Qwen3-4B-abliterated
8B source: huihui-ai/Qwen3-8B-abliterated
VL 4B source: huihui-ai/Huihui-Qwen3-VL-4B-Thinking-abliterated
Original base family: Qwen/Qwen3

Citation

If you use these models, please cite the original Qwen3 work:

bibtex

@misc{qwen3,
  title={Qwen3: A Large-Scale Multilingual Language Model},
  author={Qwen Team},
  year={2024},
  howpublished={\url{https://github.com/QwenLM/Qwen3}}
}