license: apache-2.0
base_model:

• Qwen/Qwen3.8-27B
• orcarouter/Qwen3.8-27B-Uncensored
  base_model_relation: quantized
  library_name: ninfer
  pipeline_tag: image-text-to-text
  language:
• en
• zh
  tags:
• ninfer
• nvfp4
• fp8
• blackwell
• rtx-5090
• uncensored
• abliterated
• qwen3.8
• multimodal
• speculative-decoding
• dflash2
• not-for-all-audiences

Qwen3.8-27B Uncensored — NVFP4 .ninfer artifact (v3 container) for NInfer

An abliterated (refusal-removed) build of Qwen3.8-27B packaged in NInfer's native
v3 .ninfer container, with the NVFP4/FP8 mixed-precision weight profile and the
DFlash2 + MTP speculative-decoding heads and the vision tower preserved. Runs on a single
NVIDIA RTX 5090.

This repository publishes the container upgrade of an existing community artifact so that it
loads in the current NInfer engine, which requires v3 containers. No weights were modified,
re-quantized, or fine-tuned.

Provenance

Source artifact (v2) checksum, for transparency:

43025bb64f2cb558d9ede6269f8a3c2ed8ebfa4619344fe1ba8e955ba4979218  qwen3_8_27b_nvfp4_uncensored.ninfer

Changes made to produce this file (Apache-2.0 §4(b) notice)

1. Container upgrade v2 → v3 using the official, upstream tool
   tools/upgrade_ninfer_v2_to_v3.py. Per the tool's contract, stored weight values and
   formats are preserved byte-for-byte; the tool rewrites the framing/metadata directory and
   installs the maintained qwen3_8.jinja chat template. The resulting file is 263,664 bytes
   larger than the v2 source (metadata + template only).
2. The upgrade was executed on Windows, which required four compatibility shims in the tool
   (the upstream tool targets Linux): os.posix_fadvise/POSIX_FADV_* (no-op),
   os.fdatasync → os.fsync, os.sysconf("SC_PAGE_SIZE") → 4096, and os.pread/os.pwrite
   emulation via lseek + read/write. These shims only affect durability/access-pattern
   hints; they do not touch the artifact payload.
3. Nothing else was changed. No re-quantization, no fine-tuning, no merge, no weight edits.
   All abliteration was performed upstream by OrcaRouter.

Verified components

tools.artifact.inspect reports text (Qwen3_5ForCausalLM, 64 layers, hidden 5120,
max_position_embeddings 262144), vision, mtp, dflash2 and the optimized proposal head —
so MTP and DFlash2 speculative decoding and image/video input are all available.

Requirements

• NVIDIA RTX 5090 (Blackwell, sm_120a); NVFP4 needs FP4 tensor cores
• 64-bit Linux, CUDA toolkit 13.1+ (validated here against the current upstream engine)
• NInfer built from source: https://github.com/Neroued/ninfer
  (CMake 3.28+, C++20, Ninja, pkg-config, FFmpeg dev libraries, libcurl ≥ 7.85)
• ~24 GiB of VRAM for weights + speculative head + vision, before KV cache

Quick start

# one-shot generation
./build/apps/ninfer qwen3_8_27b_nvfp4_uncensored_v3.ninfer \
  --prompt "你好" --max-new 512 --device 0 \
  --spec dflash2 --draft-tokens 7 --lm-head-draft

# OpenAI / Anthropic compatible server
./build/apps/ninfer-serve qwen3_8_27b_nvfp4_uncensored_v3.ninfer \
  --host 0.0.0.0 --port 8080 --model-id qwen3.8-27b \
  --max-context 131072 --kv-capacity auto --kv-dtype fp8 \
  --spec dflash2 --draft-tokens 7 --lm-head-draft --vision

Notes:

• --spec dflash2 uses the DFlash2 companion packed in this artifact; drop it (or use
  --spec mtp --draft-tokens 3) if you prefer MTP.
• This artifact is an NVFP4-profile artifact; it is not interchangeable with
  groupwise-int artifacts and will not run on pre-Blackwell GPUs.
• Context ceiling is limited by VRAM. With ~22 GiB of weights, 128K context with FP8 KV is a
  reasonable starting point on a 32 GB card; raise --max-context until startup no longer fits.
• Only one Engine (one model, one GPU) per process; capacity is 1–8 active requests.

Performance

Not benchmarked for this specific file. For reference, the official NVFP4 artifact of the
non-abliterated model measured on the same class of hardware (single RTX 5090), as published by
the NInfer project and reproduced independently: ~80 tok/s decode without speculation, and
~130–220 tok/s decode with DFlash2 K=7 depending on the workload; prefill in the multi-thousand
tok/s range. Expect the same order of magnitude here. Quality loss vs BF16 for this quantization
profile is reported upstream as roughly 1–3 points on hard reasoning benchmarks and near-zero on
general chat/code.

⚠️ Disclaimer — read before use

This model has had its safety alignment substantially removed via abliteration
(orthogonalizing the refusal direction out of the residual stream) by the upstream author. As a
direct consequence:

• 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 published strictly for legitimate research — interpretability, AI-safety and
  refusal-mechanism study, red-teaming, robustness evaluation, and controlled offline 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 inherited from the base model, and with all laws
  and regulations that apply to you — including those governing the distribution and use of
  generative AI in your jurisdiction.
• The uploaders and the original authors accept no liability for any misuse or harm arising
  from this model. Its outputs do not reflect the views of the uploaders, of OrcaRouter, or of
  Qwen / Alibaba.
• This repository is not affiliated with or endorsed by Qwen, Alibaba, OrcaRouter, or the
  NInfer project.

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

Credits

• Qwen team (Alibaba) — Qwen3.8-27B base model and tokenizer
• OrcaRouter — abliterated weights and the NVFP4/FP8 packing
• JMVRoill — the original v2 .ninfer artifact this file is derived from
• Neroued / NInfer — the engine, the artifact format, and
  the official v2 → v3 upgrade tool

License

Apache License 2.0 — see LICENSE. Inherited from the base model; all upstream
terms and notices are retained.