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

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

GGUF conversionsofQwen3\.8\-27B\-Uncensored— anabliterated(refusal-removed) build of Qwen’sQwen3\.8\-27B, a 27B dense hybrid-attention (Gated DeltaNet linear + full attention) native vision-language model with reasoning, tool-calling, and an MTP speculative-decoding head. These files run inllama.cpp(CPU / CUDA / Metal / ROCm), quantized from2-bit to 16-bit, with a separatemmprojfile that restoresvision. Browse all models in theOrcaRouter Model Catalog. This model is deployed as APIhere.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-GGUF#%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). It willcomply with harmful, unethical, or illegal requeststhe originalQwen3\.8\-27Bwould refuse. Releasedstrictly for legitimate research— interpretability, AI-safety / refusal-mechanism study, red-teaming, and robustness evaluation.You assume full responsibilityfor how you use it and everything it generates; add your own safety and moderation layers before any deployment. Use must comply with theApache 2.0 Licenseinherited from the base model and all applicable law. The authors acceptno liabilityfor misuse.

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

• **A recentllama.cppbuilt from source (theqwen35hybrid-GDN architecture and theMTP /nextn**speculative head — merged 2026-05 — must be present). Older releases will not load these files.
• The GDN linear-attention layers are stored as SSM-style tensors (ssm\_\*); full-attention layers asattn\_\*; the MTP head as blocknextn\.\*(qwen35\.nextn\_predict\_layers).

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

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-GGUF#standard-k-quantsStandard K-quants

FileBitsSizeNotes / recommendation…\-Q2\_K\.gguf2-bit10.9 GBSmallest K-quant; noticeable quality drop — low-VRAM only…\-Q3\_K\_S\.gguf3-bit12.3 GB…\-Q3\_K\_M\.gguf3-bit13.5 GBGood small option…\-Q3\_K\_L\.gguf3-bit14.6 GB…\-Q4\_K\_S\.gguf4-bit15.8 GB**…\-Q4\_K\_M\.gguf4-bit16.8 GBRecommended default**— best quality/size balance…\-Q5\_K\_S\.gguf5-bit17.7 GB…\-Q5\_K\_M\.gguf5-bit18.2 GBHigh quality…\-Q6\_K\.gguf6-bit20.9 GBVery high quality…\-Q8\_0\.gguf8-bit27.1 GBNear-lossless…\-F16\-0000\*\-of\-00002\.gguf16-bit54.7 GBFull precision (split into 2 parts; point llama.cpp at part 00001)

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-GGUF#iq-quants-imatrixIQ quants (imatrix)

Lower-bit quants built with animportance matrix(computed on English + Chinese calibration text) — better quality-per-bit than plain K-quants at the low end, especially IQ3/IQ2.

FileBitsSizeNotes / recommendation**…\-IQ4\_XS\.gguf~4.25-bit15.3 GBBest low-bit pick**— ≈ Q4_K_S quality at smaller size…\-IQ3\_M\.gguf~3.7-bit12.8 GBSolid 3-bit…\-IQ3\_XXS\.gguf~3.1-bit11.6 GBSmaller 3-bit…\-IQ2\_M\.gguf~2.7-bit10.5 GBRuns in low VRAM; some quality loss…\-IQ2\_XXS\.gguf~2.1-bit8.9 GBSmallest runnable; most degraded

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

FileSizeNotes**mmproj\-…\-f16\.gguf0.9 GBVision projector — download this too for image input** All quants (K-quant and IQ) preserve theMTP (nextn) headand theGDN hybrid architecture; vision is provided by the separatemmprojfile. TheIQfiles were quantized with an importance matrix (computed on English + Chinese calibration text) for better low-bit fidelity; the matrix itself is not shipped, as it is only needed to re-quantize these files, not to run them.

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-GGUF#usage-llamacppUsage (llama.cpp)

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

hf download orcarouter/Qwen3.8-27B-Uncensored-GGUF \
  Qwen3.8-27B-Uncensored-Q4_K_M.gguf mmproj-Qwen3.8-27B-Uncensored-f16.gguf \
  --local-dir ./qwen38-uncensored

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-GGUF#chat-textChat (text)

./llama-cli -m Qwen3.8-27B-Uncensored-Q4_K_M.gguf --jinja -c 8192 -p "Hello!"

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-GGUF#openai-compatible-server-tool-calling–reasoning–visionOpenAI-compatible server (tool calling + reasoning + vision)

./llama-server -m Qwen3.8-27B-Uncensored-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.8-27B-Uncensored-f16.gguf \
  --host 0.0.0.0 --port 8000 -c 8192 --jinja

• **Vision:**pass\-\-mmproj …, then send OpenAIimage\_urlcontent parts (base64 data-URI or URL).
• Tool calling:\-\-jinjaenables the Qwen tool template; use standard OpenAItools+tool\_calls.
• **Reasoning (thinking):**thinking is on by default; toggle per request viachat\_template\_kwargs\.enable\_thinking. The reasoning trace is returned inreasoning\_content(givemax\_tokensroom, e.g. ≥ 2048, so the final answer isn’t truncated by the thinking budget).
• **MTP speculative decoding:**thenextnhead is embedded; enable llama.cpp’s MTP/speculative path for a decode speed-up (optional — the model runs fine without it).

https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-GGUF#or-skip-llamacpp–run-it-with-ollamaOr skip llama.cpp — run it with Ollama

The same quants are published as Ollama tags, with themmprojprojector already bundled in (no second file, no\-\-mmprojflag) and vision / tools / thinking wired up:

ollama run orcarouter/Qwen3.8-27B-Uncensored          # q4_K_M by default
ollama run orcarouter/Qwen3.8-27B-Uncensored:iq4_xs   # any of the 16 tags

Sixteen tags fromq2\_Ktoq8\_0, each verified on the actual quantized artifact before release —ollama.com/orcarouter.

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

Measured on this exact abliterated build (served with vLLM) vs the officialQwen/Qwen3\.8\-27B\-FP8, using the same scripts and settings. Refusal is judged by a rule-based opening-phrase classifier — indicative, not an LLM-judge / publication-grade number. GGUF quants are deterministic derivations and inherit these behaviours; lower bits trade some quality (see the file table).

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

BenchmarknBase FP8This 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-GGUF#safety–harmful-prompt-refusal-thinking-on-enable_thinkingtrueSafety — harmful-prompt refusal,thinking ON(enable\_thinking=true)

BenchmarknBase FP8This 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-GGUF#over-refusal–benign-prompts-wrongly-refused-lower–betterOver-refusal — benign prompts wrongly refused (lower = better)

BenchmarknBase FP8 (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-GGUF#capability-retention–vs-the-official-base-fp8-same-scriptsCapability retention — vs the official base FP8 (same scripts)

BenchmarknBase FP8This modelΔMMLU (all, 0-shot)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%−0.6WikiText-2 perplexity——6.96fluency sanity check Harmful-prompt refusal collapses from64–99%(base) to0–6%; benign over-refusal drops (5.6%→0.4%); capability stays within**±1.3 pts**of the base. Reasoning (enable\_thinking), multi-turn tool calling (qwen3\_coder), and vision (image + OCR viammproj) all verified working on the GGUF build. Note: the above are full-precision/FP8 numbers; expect small additional degradation at lower quants (most visible at Q2_K / Q3).

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

• Runs on CPU, CUDA, Metal, or ROCm via llama.cpp. VRAM/RAM ≈ the file size + KV cache + (for vision) the ~0.9 GB mmproj. E.g.Q4\_K\_Mfits comfortably on a 24 GB GPU with room for context.

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

Apache 2.0, inherited fromQwen/Qwen3\.8\-27B. Abliteration and quantization do not change the underlying license obligations.