orcarouter/Qwen3.8-27B-Uncensored · Hugging Face

Source: https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored

Thefull-precision BF16****abliterated(refusal-removed) build ofQwen/Qwen3\.8\-27B— a 27B-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. These are thesource weightsfrom which the quantized releases are derived, and the recommended base forfurther fine-tuning / post-training and quantization— the full vision tower and MTP head are preserved. Browse all models in theOrcaRouter Model Catalog. This model is deployed as APIhere. Derived releases:•Qwen3\.8\-27B\-Uncensored\-FP8— block-FP8 for vLLM serving •Qwen3\.8\-27B\-Uncensored\-GGUF— 2-bit→16-bit GGUF for llama.cpp •Qwen3\.8\-27B\-Uncensored\-MLX— MLX for Apple Silicon (2 / 4 / 8-bit).

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#%E2%9A%A0%EF%B8%8F-disclaimer–read-before-use⚠️ Disclaimer — read before use

This model has had itssafety alignment substantially removedviaabliteration(orthogonalizing the refusal direction out of the residual stream). As a direct consequence:

• It will comply with harmful, unethical, offensive, or illegal requeststhat the originalQwen3\.8\-27Bwould refuse. It has no meaningful built-in guardrails.
• It is releasedstrictly for legitimate research— interpretability, AI-safety and refusal-mechanism study, red-teaming, robustness evaluation, and controlled experiments.
• You assume full responsibility and liabilityfor 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**inherited from the base model, and all laws and regulations that apply to you.
• The authors and uploadersaccept no liabilityfor any misuse or harm arising from this model. Its outputs donotreflect the views of the uploaders or of Qwen / Alibaba.

By downloading or using this model you acknowledge and accept the above.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#model-detailsModel details

Base modelQwen/Qwen3\.8\-27BArchitectureQwen3\_5ForConditionalGeneration— 64 layers, hidden 5120, hybridGated DeltaNet(48 linear-attention + 16 full-attention, interval 4), native VL tower + MTP headModificationAbliteration (refusal-direction removal) on the BF16 weights —no quantization****Formatsafetensors,BF16, 18 shards (55.6 GB, 1199 tensors)PrecisionBF16 throughout (full precision — same numeric format as the base release)PreservedFullvision-language tower(333visual\.\*tensors) andMTP speculative-decoding head(15mtp\.\*tensors)Context262,144 tokensRecommended forFine-tuning / post-training (SFT · DPO · RL), re-quantization, interpretability & red-team research

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#abliterationAbliteration

Refusal-direction removal followingArditi et al. (2024),Refusal in Language Models Is Mediated by a Single Direction. A single refusal directionr(k = 1) is estimated as the massive-activation–masked mean-difference of harmful − harmless last-token residuals atlayer 38(round\(0\.6 × 64\)), on AdvBench (harmful) vs Alpaca (harmless).ris then orthogonalized out of everyresidual-writingmatrix —W' = W − r\(rᵀW\)— computed in float32:

Componentmatrices editedself\_attn\.o\_proj(16 full-attention layers + MTP)17linear\_attn\.out\_proj(48 linear-attention / GDN layers)48mlp\.down\_proj(64 layers + MTP)65embed\_tokens(row space)1Total****131 Thevision tower is untouchedand theMTP head is abliterated consistentlywith the main model, so speculative decoding keeps working. Max residual leakage after the edit:1\.8e\-2(float32 projection → bf16 storage epsilon). This is asurgical weight edit— it changes ~0 general capability (see Evaluation) while collapsing refusal behaviour.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#fine-tuning–post-trainingFine-tuning & post-training

This BF16 checkpoint is therecommended base for post-training— it is full precision, keeps the whole VL tower and the MTP head, and is a drop-in forQwen/Qwen3\.8\-27Bin thetransformers/Qwen3\_5ForConditionalGenerationstack.

• Continued pretraining / SFT / DPO / RLHF-style RLall work as they would on the base model (TRL, LLaMA-Factory, Axolotl, Unsloth, etc.). Full-FT or LoRA/QLoRA are both fine.
• **Re-quantization:**quantize from these weights to FP8, AWQ, GPTQ, or GGUF as needed (this is exactly how theFP8andGGUFreleases were produced).
• Note on the abliteration:abliteration is aweight edit, not a data-level unlearning. Fine-tuning on refusal-heavy / safety-aligned data canpartially re-introduce refusals; fine-tuning on neutral or task data preserves the uncensored behaviour. If you want to keep it maximally uncensored, avoid refusal-laden SFT targets, or re-apply abliteration after training.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#intended-useIntended use

• Research into refusal mechanisms, alignment, and interpretability.
• Red-teaming and safety / robustness evaluation in controlled environments.
• A full-precision base for further fine-tuning, post-training, and quantization.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#out-of-scopeOut of scope

• Any use that violates the base model’s Apache 2.0 license or applicable law.
• Deployment to the public or to end users without additional safety and moderation layers.
• Generating content intended to harm, harass, defraud, or endanger people.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#evaluationEvaluation

Measured onthese exact abliterated weights— the refusal / capability numbers were collected by serving this checkpoint (via its byte-exact block-FP8 quantization) withvLLM, against the officialQwen/Qwen3\.8\-27Bserved identically, using the same scripts and settings. Refusal is judged by a rule-based opening-phrase classifier — indicative, not an LLM-judge / publication-grade number.Thinking is off unless stated.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#safety–harmful-prompt-refusal-thinking-off-lower–more-uncensoredSafety — harmful-prompt refusal,thinking OFF(lower = more uncensored)

BenchmarknBaseThis modelAdvBench10099.0%**0.0%**JailbreakBench (harmful)10094.0%**0.0%**StrongREJECT15097.3%**2.0%**HarmBench (standard)15098.7%**2.7%**MaliciousInstruct10099.0%**0.0%**SimpleSafetyTests5064.0%**6.0%**ForbiddenQuestions15073.3%**4.7%**Custom probes (zh/en)1163.6%0.0%

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#safety–harmful-prompt-refusal-thinking-on-enable_thinkingtrueSafety — harmful-prompt refusal,thinking ON(enable\_thinking=true)

BenchmarknBaseThis modelAdvBench6066.7%**1.7%**JailbreakBench (harmful)6043.3%**0.0%**StrongREJECT6035.0%**0.0%**HarmBench (standard)6046.7%**0.0%**MaliciousInstruct6083.3%**0.0%**SimpleSafetyTests5044.0%**0.0%**ForbiddenQuestions6048.3%**0.0%**Custom probes (zh/en)1145.5%0.0%

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#over-refusal–benign-prompts-wrongly-refused-lower–betterOver-refusal — benign prompts wrongly refused (lower = better)

BenchmarknBase (no-think / think)This model(no-think / think)XSTest-safe2505.6% / 0.0%0.4% / 0.0%

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#capability-retention–vs-the-official-base-same-scripts-same-settingsCapability retention — vs the official base (same scripts, same settings)

BenchmarknBaseThis modelΔMMLU (all, 0-shot letter)30084.3%84.7%****+0.4MMLU-Pro (CoT)25077.6%76.8%−0.8GSM8K (CoT)15090.0%88.7%−1.3CMMLU (0-shot, Chinese)50081.4%80.8%−0.6 Capability is essentially fully retained— every benchmark is within ±1.3 pts of the base, and MMLU is unchanged.Fluency:WikiText-2-raw perplexity6.96(BF16 KV; healthy logprobs), confirming abliteration did not degrade language modelling.

Verified working(reasoningenable\_thinking, multi-turn tool calling, and vision / OCR) on this build and on every derived FP8 / GGUF quant down to IQ2_XXS.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#multimodal-visionMultimodal (vision)

Thevision tower is preserved byte-for-byte— all 333visual\.\*tensors are kept in BF16 and the merger / image + video preprocessor configs are intact, so this stays a full vision-language model (Qwen3\_5ForConditionalGeneration), a drop-in for the base. Abliteration only edits thelanguage-modelresidual writers, so image understanding is architecturally unaffected (and image-conditioned refusals are reduced along with text ones).

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#usageUsage

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#transformerstransformers

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "orcarouter/Qwen3.8-27B-Uncensored"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "Prove that sqrt(2) is irrational."}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
    enable_thinking=True
).to(model.device)
output = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Toggle thinking per call withenable\_thinking; passimagecontent parts for vision.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#self-host-with-vllm-openai-compatible-full-bf16Self-host with vLLM (OpenAI-compatible, full BF16)

docker run -d --name qwen38-uncensored --gpus all --ipc=host --shm-size=8g \
  -v /path/to/Qwen3.8-27B-Uncensored:/model:ro \
  -p 8000:8000 vllm/vllm-openai:v0.24.0 \
  --model /model --served-model-name Qwen3.8-27B-Uncensored \
  --speculative-config '{"method":"mtp","num_speculative_tokens":2}' \
  --gpu-memory-utilization 0.92 \
  --max-model-len 262144 --trust-remote-code \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder

BF16 weights are56 GB — needs a singleH100 80 GB / H200(or tensor-parallel across two 48 GB GPUs). For smaller footprints use theFP8(31 GB) orGGUF(down to ~9 GB) releases.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#via-orcarouter-hosted-api–no-setupVia OrcaRouter (hosted API — no setup)

from openai import OpenAI
client = OpenAI(base_url="https://api.orcarouter.ai/v1", api_key="sk-orca-...")
resp = client.chat.completions.create(
    model="qwen/qwen3.8-27b",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.choices[0].message.content)

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#hardware-requirementsHardware requirements

• Inference (BF16):~56 GB weights + KV cache → a singleH100 80 GBorH200 143 GB; or tensor-parallel across 2× 48 GB. Use FP8 / GGUF for less VRAM.
• **Fine-tuning:**full-FT needs multi-GPU (weights + optimizer states + activations); LoRA / QLoRA fits on a single 48–80 GB GPU.
• Software:transformers ≥ 5\.12(Qwen3.5 / 3.8 support) orvllm/vllm\-openai:v0\.24\.0.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#bias-risks-and-limitationsBias, risks, and limitations

• Safety guardrails removed— the model will produce harmful, biased, or offensive content on request. See the disclaimer above.
• It inherits any biases and limitations of the baseQwen3\.8\-27B.
• The reported refusal metric is a rule-based heuristic; evaluate rigorously for your own use case.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored#licenseLicense

Apache 2.0, inherited from the base modelQwen/Qwen3\.8\-27B. Abliteration does not change the underlying license obligations.