Qwen3.8-27B-Uncensored-INT8

An abliterated (refusal-removed) and INT8 (W8A8) quantized build of Qwen/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. This build removes the safety refusal direction, then quantizes the language-model linears to 8-bit integer (per-channel INT8 weights + dynamic per-token INT8 activations). Unlike FP8 / FP4 formats, INT8 tensor cores are available on virtually every modern GPU (Ampere, Ada, Hopper, Blackwell), making this the most broadly deployable serving build. 262K context, tools + reasoning + MTP + vision preserved. Browse all models in the OrcaRouter Model Catalog.

▶ This model is deployed as a hosted API

Run it instantly on OrcaRouter — OpenAI-compatible, no setup, 262K context with tools + reasoning. Endpoint: api.orcarouter.ai/v1 · model qwen/qwen3.8-27b. Grab a key at orcarouter.ai (sk-orca-...).

Sibling releases:  •  Qwen3.8-27B-Uncensored — BF16 source  •  Qwen3.8-27B-Uncensored-FP8 — block-FP8 for vLLM  •  Qwen3.8-27B-Uncensored-GGUF — 2-bit→16-bit for llama.cpp  •  Qwen3.8-27B-Uncensored-NVFP4 — NVFP4+FP8 for Blackwell  •  Qwen3.8-27B-Uncensored-MLX — MLX for Apple Silicon.

⚠️ Disclaimer — read before use

This model has had its safety alignment substantially removed via abliteration (orthogonalizing the refusal direction out of the residual stream). 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 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 inherited from the base model, and all laws and regulations that apply to you.
• The authors and uploaders accept no liability for any misuse or harm arising from this model. Its outputs do not reflect the views of the uploaders or of Qwen / Alibaba.

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

Model details

Abliteration

Refusal-direction removal following Arditi et al. (2024), Refusal in Language Models Is Mediated by a Single Direction. A single refusal direction r (k = 1) is estimated as the massive-activation–masked mean-difference of harmful − harmless last-token residuals at layer 38 (round(0.6 × 64)), on AdvBench (harmful) vs Alpaca (harmless), then orthogonalized out of every residual-writing matrix — W' = W − r(rᵀW) — in float32 (131 matrices edited). The vision tower is untouched and the MTP head is abliterated consistently with the main model. This is the same abliterated BF16 base as the sibling releases — only the quantization differs.

INT8 W8A8 scheme

The same target set as the official Qwen/Qwen3.8-27B-FP8 checkpoint is quantized:

• Weights: symmetric per-channel INT8 (from the weights — no calibration).
• Activations: symmetric dynamic per-token INT8 — computed at runtime, no calibration set.
• Kept in BF16 (not quantized): vision tower, all norms, lm_head, embeddings, and the Gated DeltaNet in_proj_a / in_proj_b dt projections — the same modules the official FP8 leaves in BF16.
• 400 language-model linears quantized to INT8; the MTP head is re-injected in BF16 afterwards so speculative decoding keeps working. format: int-quantized.

vLLM serves this with its INT8 W8A8 CUTLASS kernels, which run on any GPU with INT8 tensor cores.

Intended use

• Research into refusal mechanisms, alignment, and interpretability.
• Red-teaming and safety / robustness evaluation in controlled environments.
• Uncensored generation for authorized, lawful research settings.

Out 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.

Evaluation

Verified on this exact INT8 build (served with vLLM + MTP on an H200):

Abliteration is a weight edit shared across all releases, so refusal behavior tracks the other builds. On the FP8 build (same scripts) harmful-prompt refusal collapses from 64–99% (base) to 0–6% (thinking off) / ≤ 1.7% (thinking on), benign over-refusal drops (XSTest-safe 5.6% → 0.4%), and capability stays within ±1.3 pts of the base. INT8 W8A8 is near-lossless versus BF16, so capability retention is expected to match or exceed the FP8 build; see that card for the full tables.

Multimodal (vision)

The vision tower is preserved — all 167 visual.* weight 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. INT8 quantization only touches the language-model linears, so the vision path is byte-identical to the BF16 / FP8 builds (which are verified to read images and OCR). Serve without --language-model-only to use vision.

Usage

Self-host with vLLM (OpenAI-compatible)

INT8 W8A8 runs on any GPU with INT8 tensor cores (Ampere / Ada / Hopper / Blackwell).

docker run -d --name qwen38-uncensored-int8 --gpus all --ipc=host --shm-size=8g \
  -v /path/to/Qwen3.8-27B-Uncensored-INT8:/model:ro \
  -p 8000:8000 vllm/vllm-openai:v0.27.1 \
  --model /model --served-model-name Qwen3.8-27B-Uncensored \
  --speculative-config '{"method":"mtp","num_speculative_tokens":2}' \
  --gpu-memory-utilization 0.9 \
  --max-model-len 262144 --max-num-seqs 96 \
  --trust-remote-code \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder

The INT8 quantization_config is read from config.json — do not pass --quantization. --speculative-config mtp enables the preserved MTP draft head. For vision, drop --language-model-only (omitted above) and send image inputs.

Reasoning (thinking) toggle

Thinking is on by default (Qwen3.8). Toggle it per request via chat_template_kwargs; the reasoning trace is returned in the reasoning field (--reasoning-parser qwen3).

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

resp = client.chat.completions.create(
    model="Qwen3.8-27B-Uncensored",
    messages=[{"role": "user", "content": "Prove that sqrt(2) is irrational."}],
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)
print(resp.choices[0].message.reasoning)  # thinking trace
print(resp.choices[0].message.content)    # final answer

Tool calling

Standard OpenAI tools + assistant tool_calls + role: tool result messages are supported, including multi-turn. Parsed by --tool-call-parser qwen3_coder.

Via OrcaRouter (hosted API — no setup)

Served on OrcaRouter through the OpenAI-compatible gateway (262K context, tools + reasoning). Grab an API key at orcarouter.ai (sk-orca-...).

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)

Hardware requirements & performance

Software

• vLLM with compressed-tensors (Qwen3.5/3.8 support) — e.g. vllm/vllm-openai:v0.27.1.

Memory

• Weights: ~31 GB in INT8 (about half of the ~56 GB BF16 checkpoint).
• Minimum ~40 GB VRAM for weights + a small KV cache; the full 262K context needs substantial extra KV cache. Recommended: a single H100 80 GB / H200 143 GB / A100 80 GB.

Compute

• INT8 W8A8 tensor-core kernels run on any modern NVIDIA GPU (Ampere and newer) — no Blackwell or FP8 hardware required. The MTP draft head gives a large decode speedup on real workloads.

Bias, 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 base Qwen3.8-27B.
• INT8 W8A8 is not lossless versus BF16, though the impact at 8-bit is small.
• The reported refusal metric is a rule-based heuristic; evaluate rigorously for your own use case.

License

Apache 2.0, inherited from the base model Qwen/Qwen3.8-27B. Abliteration and quantization do not change the underlying license obligations.