Qwen3.8-27B-Uncensored-FP8: What We Shipped, and Why It's Research-Only

We shipped Qwen3.8-27B-Uncensored-FP8 today — an abliterated, offline block-FP8 build of Qwe​n's Qwen3.8 27B, hosted at orcarouter/Qwen3.8-27B-Uncensored-FP8 on Hugging Face. This post is the honest version of what it is, what we did, and — most importantly — what it is not for.

What this model is

Qwen3.8 27B is a 27-billion-parameter dense, hybrid-attention (Gated DeltaNet linear + full attention) native vision-language model with flexible thinking control, tool-calling, and an MTP speculative-decoding head. The FP8 build re-quantizes it to the exact FP8 scheme of the official Qwen3.8-27B-FP8, so it serves on the identical vLLM kernel path — 262K context, tools, reasoning, and MTP all preserved.

What makes this build different is the word before FP8: uncensored. We applied abliteration — orthogonalizing the refusal direction out of the residual stream — to remove the model's safety-alignment refusals. The result is a model that will comply with requests the original would refuse.

Two key artifacts:

• Hugging Face repo — orcarouter/Qwen3.8-27B-Uncensored-FP8 — safetensors, Apache 2.0, block-FP8 E4M3, image-text-to-text.

• Model card on OrcaRouter — live under our Obsidian catalog line, with pricing and benchmarks.

Why we shipped it

Because refusal mechanisms are a legitimate research subject. Interpretability teams study how refusal directions are encoded in the residual stream. Safety researchers probe what happens when alignment is removed. Red-teaming requires models that don't refuse, so you can test your own guardrails against something that actively tries to bypass them. Abliterated models are a standard tool in that kit.

We built this on Qwe​n's open weights (Apache 2.0) and re-released it under the same license. The point is to give researchers a known, reproducible abliterated baseline — the same architecture, the same FP8 serving path, one variable changed.

The part you must read before downloading

This model has had its safety alignment substantially removed. Concretely:

• It will comply with harmful, unethical, offensive, or illegal requests that the original Qwen3.8 27B would refuse. It has no meaningful built-in guardrails.

• It is released strictly for legitimate research — interpretability, AI-safety and refusal-mechanism study, red-teaming, robustness evaluation, and controlled experiments.

• You assume full responsibility and liability for how you use it and for everything it generates. Do not deploy it to end users or in production without adding your own safety, moderation, and abuse-prevention layers.

• Use must comply with the Apache 2.0 License.

If your use case is a production product, a chatbot, or anything that serves end users — this is not the model for you. Use the standard Qwen3.8 27B, which keeps its safety alignment.

How to use it

If you're doing legitimate research and understand the above:

• Pull it from Hugging Face: orcarouter/Qwen3.8-27B-Uncensored-FP8

• Or access it through our model card, which has the pricing and benchmark details.

• It serves on the standard vLLM kernel path with 262K context, tools, reasoning, and MTP intact.

We ship this because refusal-removal research is real and underserved. We also ship it with a disclaimer that doesn't hide what it is. Read it, understand it, and use it within the bounds it's meant for.

For legitimate research, the weights are on Hugging Face: Pull it from Hugging Face