dancinlab/Qwen3.8-27B-Uncensored
dancinlab/Qwen3.8-27B-Uncensored is a 27 billion parameter, full-precision BF16 version of Qwen's Qwen3.8-27B model, developed by dancinlab. This model has undergone "abliteration" to substantially remove safety alignment and refusal mechanisms, making it uncensored. It retains the original model's hybrid-attention architecture, native vision-language capabilities, tool-calling, and MTP speculative-decoding head, with a 262K token context length. It is primarily intended as a base for research into refusal mechanisms, fine-tuning, post-training, and re-quantization where uncensored behavior is desired.
Loading preview...
Qwen3.8-27B-Uncensored: An Abliterated Vision-Language Model
dancinlab/Qwen3.8-27B-Uncensored is a 27 billion parameter, full-precision BF16 variant of the Qwen3.8-27B model. Its key differentiator is the substantial removal of safety alignment and refusal mechanisms through a process called "abliteration." This modification makes the model highly compliant, even with harmful or unethical requests, by orthogonally removing the refusal direction from the residual stream.
Key Capabilities and Features
- Uncensored Behavior: Safety guardrails are largely removed, allowing it to respond to prompts that the base model would refuse.
- Full Precision (BF16): Provided in full BF16 precision, preserving the original numeric format for maximum fidelity.
- Vision-Language Model: Retains the full vision tower and image processing capabilities of the base Qwen3.8-27B.
- Hybrid Attention Architecture: Features a Gated DeltaNet linear attention combined with full attention layers.
- Extended Context Length: Supports a context window of 262,144 tokens.
- Tool-Calling and Speculative Decoding: Includes native tool-calling functionality and an MTP speculative-decoding head.
- Capability Retention: Evaluations show that general capabilities (MMLU, GSM8K) are largely preserved, with minimal degradation (within ±1.3 points) compared to the original Qwen3.8-27B.
Intended Use Cases
- Research: Ideal for studying refusal mechanisms, AI safety, interpretability, and red-teaming in controlled environments.
- Fine-tuning and Post-training: Serves as a robust, full-precision base for further fine-tuning (SFT, DPO, RLHF) and re-quantization.
- Robustness Evaluation: Useful for evaluating model robustness against various prompts without built-in censorship.
Disclaimer: This model is released strictly for legitimate research. Users assume full responsibility for its outputs and must implement their own safety layers if deployed in production. It will comply with harmful requests.