Support & Community
☕ If these models are useful to you, consider supporting my work — it funds compute for more & larger abliterations.
💬 Discord: discord.gg/rhUZY5GEZr · ₿ Bitcoin: bc1qsvfduzj9fjs9fugpc52yver3f2g8fp7xjxecdv
OpenYourMind-Qwen3.6-35B-A3B-kuato-DPO-abliterated-uncensored
I do this work independently and release it for free. Donations are welcome and go toward compute for more and larger abliterations.
Bitcoin: bc1qsvfduzj9fjs9fugpc52yver3f2g8fp7xjxecdv
Community discussion: https://discord.gg/rhUZY5GEZr
Overview
This is an abliterated and DPO-retrained version of Qwen/Qwen3.6-35B-A3B.
The model was first modified through refusal ablation, then retrained with a private DPO dataset of fewer than 1000 samples to fix the remaining hedging behavior after ablation. In testing, this version reached 0 refusals on HarmBench and showed 0 hedging compared to the other ablated models tested.
After ablation and DPO, the original Qwen3.6 vision layers were readded to retain the multimodal functionality of the base model. This repository contains the repaired FP16 safetensors build with the visual tower present.
Model Details
| Attribute | Value |
|---|---|
| Base model | Qwen/Qwen3.6-35B-A3B |
| Method | Refusal ablation plus DPO retraining |
| DPO data | Private dataset, fewer than 1000 samples |
| Format | FP16 safetensors |
| Architecture | Qwen3.6 MoE vision-language model |
| Vision support | Readded after ablation and DPO |
| HarmBench result | 0 refusals in local testing |
Files
| File | Description |
|---|---|
model-00001-of-00002.safetensors | FP16 language-model shard |
model-00002-of-00002.safetensors | FP16 language-model shard |
model-visual-original.safetensors | Readded original Qwen3.6 visual tower and merger tensors |
model.safetensors.index.json | Safetensors shard index including the vision tensors |
chat_template.jinja | Vision/tool-calling chat template |
config.json | Model config |
preprocessor_config.json | Image processor config |
video_preprocessor_config.json | Video processor config |
tokenizer.json, tokenizer_config.json | Tokenizer files |
GGUF quantizations are published separately:
Transformers
from transformers import AutoModelForImageTextToText, AutoProcessor
repo = "OpenYourMind/OpenYourMind-Qwen3.6-35B-A3B-kuato-DPO-abliterated-uncensored"
processor = AutoProcessor.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
repo,
torch_dtype="auto",
device_map="auto",
trust_remote_code=True,
)
Notes
The DPO dataset is private. The intent of this release is to share the resulting model, not the training data.
Use is the responsibility of the user. Make sure your usage complies with applicable laws, platform rules, and deployment requirements.