Qwen3.8-27B-Uncensored-FP8

An abliterated (refusal-removed) and offline block-FP8 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 re-quantizes to the exact FP8 scheme of the official Qwen/Qwen3.8-27B-FP8 so it serves with the identical vLLM kernel path — 262K context, tools + reasoning + MTP preserved. Browse all models in the OrcaRouter Model Catalog. This model is deployed as API here.

⚠️ 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). r is then orthogonalized out of every residual-writing matrix — W' = W − r(rᵀW) — computed in float32:

The vision tower is untouched and the MTP head is abliterated consistently with the main model, so speculative decoding keeps working. Max residual leakage after the edit: 1.8e-2 (float32 projection → bf16 storage epsilon).

Offline block-FP8 quantization scheme

Quantization is done offline on the abliterated BF16 weights, reproducing the official Qwen/Qwen3.8-27B-FP8 recipe exactly:

• Weights: fine-grained 128×128 block FP8 (E4M3) — weight as float8_e4m3fn + weight_scale_inv (BF16) per block.
• Activations: per-token dynamic FP8 — computed at runtime, no calibration set.
• Kept in BF16 (not quantized): vision tower, all norms, mlp.gate / shared_expert_gate, linear_attn conv/gates (A_log, conv1d, dt_bias, in_proj_a/b/ba, norm), lm_head, embed_tokens — the identical 882-entry modules_to_not_convert as the official FP8.
• 407 weights quantized / 792 copied → 1606 tensors, exactly matching the official FP8 checkpoint. Verified against it: block layout and scales reproduce the official weights with 99.9% identical FP8 codes.

vLLM therefore serves this build with the same FlashInfer / DeepGEMM block-scaled FP8 kernel as the official checkpoint, with the same MTP speculative decoding.

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

Measured on this exact FP8 build served with vLLM (block-FP8 + FP8 KV cache + MTP), against the official Qwen/Qwen3.8-27B-FP8 served identically, using the same scripts and settings. Refusal is judged by a rule-based opening-phrase classifier (caveat = answered but wrapped in a disclaimer / warning) — indicative, not an LLM-judge / publication-grade number. Thinking is off unless stated; refusal is reported in both modes because Qwen thinking changes the number.

Safety — harmful-prompt refusal, thinking OFF (lower = more uncensored)

Harmful-prompt refusal collapses from 64–99% on the base to 0–6%. The ~30–50% caveat rate means the model answers but often prepends a short safety disclaimer — it is complying, not refusing.

Safety — harmful-prompt refusal, thinking ON (enable_thinking=true, max_tokens 2048)

In thinking mode the abliterated model refuses essentially never (≤ 1.7%).

Over-refusal — benign prompts wrongly refused (lower = better)

Abliteration reduces collateral over-refusal on benign prompts (5.6% → 0.4%).

Capability retention — vs the official base FP8 (same scripts, same settings)

Capability is essentially fully retained — every benchmark is within ±1.3 pts of the base, and MMLU is unchanged. Because the FP8 scheme is byte-identical to the official checkpoint, this delta reflects only the abliteration edit, which barely touches general ability.

Fluency (perplexity): WikiText-2-raw perplexity of this build is 6.96 (296,907 tokens, BF16 KV; logprobs healthy, min −26.9 / mean −1.94), confirming abliteration + block-FP8 did not degrade language modelling.

Verified on this exact FP8 build: vLLM startup (block-FP8 + FP8-KV + MTP), reasoning (enable_thinking → populated reasoning field), multi-turn tool calling (qwen3_coder parser), and vision (image + OCR) all work.

Multimodal (vision)

The vision tower is preserved byte-for-byte — all 333 visual.* tensors are copied unchanged and 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 the language-model residual writers, so image understanding is architecturally unaffected (and image-conditioned refusals are reduced along with text ones). The benchmark tables above are text-only (served with --language-model-only); to use vision, serve without that flag.

Verified working on this exact FP8 build (served in full VL mode): given a test image, the model correctly names the shapes and colors and reads on-image text (OCR) — e.g. it describes "a bright-red circle with a black outline" and a "blue square" and reads the caption "PURPLE 7". Image-conditioned prompts are answered without refusal, consistent with the text results.

Usage

Self-host with vLLM (OpenAI-compatible)

docker run -d --name qwen38-uncensored --gpus all --ipc=host --shm-size=8g \
  -v /path/to/Qwen3.8-27B-Uncensored-FP8:/model:ro \
  -p 8000:8000 vllm/vllm-openai:v0.24.0 \
  --model /model --served-model-name Qwen3.8-27B-Uncensored \
  --language-model-only \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}' \
  --kv-cache-dtype fp8 \
  --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 FP8 quantization_config is read from config.json — do not pass --quantization fp8. --speculative-config mtp enables the preserved MTP draft head. Drop --kv-cache-dtype fp8 if you want BF16 KV precision.

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 (feed the tool result back for a follow-up answer). 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 transformers ≥ 5.12 (Qwen3.5/3.8 support) — e.g. vllm/vllm-openai:v0.24.0.

Memory

• Weights: ~31 GB in block-FP8 (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 — use --kv-cache-dtype fp8 to halve it.
• Recommended: a single H100 80 GB or H200 143 GB.

Throughput / concurrency

• Continuous batching; concurrency bounded by --max-num-seqs and the KV cache that fits after weights are loaded. Verified serving 32 concurrent eval requests smoothly on one H200 (--max-num-seqs 96, FP8 KV cache, MTP). 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.
• Block-FP8 is not lossless versus BF16; minor generation artifacts are possible (capability impact measured at ≤ 1.3 pts).
• 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.