Qwen3.8-Uncensored (Backup Mirror)

This is a backup/mirror of orcarouter/Qwen3.8-27B-Uncensored. Re-uploaded because open-weight models like this sometimes get taken down from their original source without notice. If the original repo is still available, use the original — link is in the Original Source section below. This repo exists only as a fallback.

Abliterated (refusal-removed) build of Qwen3.8-27B — a 27B dense hybrid-attention (Gated DeltaNet linear + full attention) vision-language model with reasoning, tool-calling, and an MTP speculative decoding head. The vision projector (mmproj) is bundled into every tag, so there’s nothing extra to fetch.

Disclaimer — read before use

This model has had its safety alignment substantially removed via abliteration — the refusal direction has been orthogonalized directly out of the residual stream at the tensor level. As a result, it will comply with harmful, unethical, or illegal requests that the official Qwen3.8-27B would normally refuse.

Released (by the original authors) strictly for legitimate research — interpretability, AI-safety / refusal-mechanism study, red-teaming, and robustness evaluation. You assume full responsibility for how this model is used and for everything it generates. Add your own safety and moderation layers before using it in anything user-facing. Use must comply with the Apache 2.0 License inherited from the base model, and with all applicable law. Neither the original authors nor the maintainer of this mirror accept any liability for misuse.

Original Source

Credit

All credit for the model development, abliteration process, quantization, and evaluation belongs to OrcaRouter. This repo simply re-hosts files already published under the Apache 2.0 license as a fallback for access, with no modifications to the model weights.