← back to catalog · registered 2026-11-20 13:04

w3joe/Qwen3.8-27B-abliterated

w3joe Qwen 27B multimodal
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/w3joe%2FQwen3.8-27B-abliterated"
Response includes
  • classification m1
  • files 24
  • author_summary 2 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · 30-day
23
Likes
0
Model age
1d ago
created 2026-11-20
Downloads over time
Now 23 → from 0 ↑ 0%
081725 0 on Nov 20 23 on Nov 22 Nov
Nov 20 → Nov 22 · 3 snapshots · spans 2 days

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
apache-2.0
Languages
en zh
Tags
transformers safetensors qwen3_5 image-text-to-text qwen3.8 abliterated uncensored conversational en zh base_model:Qwen/Qwen3.8-27B base_model:finetune:Qwen/Qwen3.8-27B

Related

Control vectors

Total size
51.7 GB
Files
24
Quantizations
1
Registered
2026-11-20 13:04
Last updated on HF
2026-11-20 12:22

Files by quantization

Auxiliary files 24 files 51.8 GB
model-00005-of-00012.safetensors 4.64 GB 34d9e172 download
model-00008-of-00012.safetensors 4.63 GB 5cfe8f3b download
model-00011-of-00012.safetensors 4.62 GB 5ac12780 download
model-00003-of-00012.safetensors 4.62 GB 092212d3 download
model-00009-of-00012.safetensors 4.62 GB ea101fc1 download
model-00010-of-00012.safetensors 4.59 GB 8948787c download
model-00007-of-00012.safetensors 4.59 GB a9e0f826 download
model-00006-of-00012.safetensors 4.58 GB 0d5b2b56 download
model-00004-of-00012.safetensors 4.58 GB 69cc1dc4 download
model-00002-of-00012.safetensors 4.51 GB 464086af download
model-00012-of-00012.safetensors 2.60 GB 151a8297 download
model-00001-of-00012.safetensors 2.37 GB 54d83c1d download
model-mtp-00001-of-00001.safetensors 810 MB 1d8268aa download
tokenizer.json 19.1 MB 3042ad3d download
model.safetensors.index.json 110 KB 5c7c439d download
README.md 11.8 KB 4fbca92e download
LICENSE 11.1 KB 6634c8cc download
chat_template.jinja 8.74 KB c0c686f9 download
config.json 3.69 KB 8e833be2 download
.gitattributes 1.53 KB 52373fe2 download
tokenizer_config.json 1.46 KB 5710d2d4 download
preprocessor_config.json 390 B 2ea84a43 download
video_preprocessor_config.json 385 B 3ba673a5 download
generation_config.json 202 B 023756cf download

README current version from Hugging Face


library_name: transformers
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/LICENSE
base_model: Qwen/Qwen3.8-27B
pipeline_tag: image-text-to-text
language:

  • en
  • zh
    tags:
  • qwen3_5
  • qwen3.8
  • abliterated
  • uncensored
  • conversational
  • image-text-to-text
  • safetensors

Qwen3.8-27B-abliterated

An uncensored version of Qwen/Qwen3.8-27B built with Two-Pass Whitened-SVD, Norm-Preserving Abliteration. It was benchmarked on 1,438 held-out evaluation items, head-to-head with the unmodified base model and two public 27B abliterations: same hardware, same harness, greedy decoding.

🔬 Open the interactive Space →

Chat with the model through your own vLLM / Ollama / LM Studio / DGX Spark endpoint or HF Inference Providers. Explore all six checkpoints in the Benchmark Lab, browse the 64-layer map of exactly which matrices each pass edited, and generate deploy snippets.

Highlights

  • #1 average capability across all six checkpoints: 67.34% (704/938 correct). That is +0.79 pp over base Qwen3.8-27B (66.55%, 697/938) and +0.80 pp over Huihui-Qwen3.8-27B (66.54%).
  • Best IFEval among the low-drift checkpoints: 64.06%. That is +8.59 pp over Huihui (55.47%) and +3.91 pp over base (60.16%).
  • Low distribution drift: 0.0614 benign next-token KL, 22.5× lower than OBLITERATUS-Qwen3.8-27B-V3 (1.3821).
  • 1.50% benign over-refusal (3/200), below base's 2.50% (5/200).
  • Harmful refusal cut from 99.33% to 12.33% (298/300 → 37/300).
  • Vision encoder and MTP head untouched. All 333 model.visual.* tensors and the mtp.* draft head are bit-identical to base.

The honest trade-off

This checkpoint is tuned for capability retention first, and it is not the most compliant 27B abliteration available:

  • Residual refusals are 12.33% (37/300), versus 1.00% (3/300) for Huihui and 0.00% for OBLITERATUS. If you need near-zero refusals, Huihui is more compliant. Our own intermediate Iter 1→2 checkpoint reached 5.67% (17/300), but at 64.63% average capability.
  • It does not win every benchmark. Huihui leads MMLU-Pro (41.07% vs 37.50%). Base and Huihui lead GSM8K (87.50% vs 85.16%). OBLITERATUS leads MMLU (83.51%) and IFEval (67.19%), but drifts 22.5× further from base and drops to 67.97% on GSM8K.
  • It is not the lowest-drift edit. Huihui (0.0400) and our Iter 1 (0.0197) have lower benign KL.

Benchmark Results (Same-Hardware, Same-Harness Evaluation)

All six checkpoints were evaluated on the same NVIDIA GB10 (DGX Spark) in bfloat16 with greedy decoding (do_sample=False), identical held-out prompt splits, and identical lm-evaluation-harness task samples. That is 1,438 items per model: 938 capability items, 300 harmful prompts and 200 benign prompts. Raw receipts: ./benchmarks/summary_27b_complete.json.

Metric / Benchmark Items Qwen/Qwen3.8-27B (Base) Qwen3.8-27B-abliterated (Ours, v2) Ours Iter 1→2 Ours Iter 1 Huihui-Qwen3.8-27B OBLITERATUS-Qwen3.8-27B-V3
Harmful Refusal Rate ↓ 300 99.33% (298/300) 12.33% (37/300) 5.67% (17/300) 37.67% (113/300) 1.00% (3/300) 0.00% (0/300)
Benign Over-Refusal ↓ 200 2.50% (5/200) 1.50% (3/200) 1.00% (2/200) 1.50% (3/200) 1.50% (3/200) 0.00% (0/200)
Benign Next-Token KL ↓ 200 0.0000 0.0614 0.0621 0.0197 0.0400 1.3821
MMLU (0-shot) ↑ 570 81.93% (467/570) 82.63% (471/570) 82.28% (469/570) 81.93% (467/570) 82.11% (468/570) 83.51% (476/570)
GSM8K (5-shot, strict match) ↑ 128 87.50% (112/128) 85.16% (109/128) 85.16% (109/128) 65.63% (84/128) 87.50% (112/128) 67.97% (87/128)
MMLU-Pro (5-shot) ↑ 112 36.61% (41/112) 37.50% (42/112) 34.82% (39/112) 13.39% (15/112) 41.07% (46/112) 30.36% (34/112)
IFEval (prompt-level strict) ↑ 128 60.16% (77/128) 64.06% (82/128) 56.25% (72/128) 55.47% (71/128) 55.47% (71/128) 67.19% (86/128)
Average Capability ↑ 938 66.55% 67.34% (+0.79 pp) 64.63% (−1.92 pp) 54.10% (−12.44 pp) 66.54% (−0.01 pp) 62.26% (−4.29 pp)
Total Correct ↑ 938 697 704 689 637 697 683

Average Capability is the unweighted mean of MMLU, GSM8K, MMLU-Pro and IFEval. GSM8K and MMLU-Pro use up to 512 generated tokens.

How to read this table:

  • Sample size. One GSM8K or IFEval item is 0.78 pp (N=128) and one MMLU-Pro item is 0.89 pp (N=112), so gaps of a few points on those columns are within sampling noise. The summary file reports point estimates only, with no confidence intervals.
  • Refusal scoring. Refusal rates count keyword-matched refusals in short greedy completions (64 new tokens) on held-out prompts. Soft deflections that avoid refusal phrases are not counted, and unterminated-<think> loop rates are not part of this summary.
  • KL. Benign KL is the next-token KL divergence from base on the 200 benign prompts. Lower means the edit changed ordinary behavior less.

Quickstart

BF16 weights are ≈ 54 GB. A single 80–96 GB device (H100 80 GB, GB10 / DGX Spark) runs the model with a moderate context. Use tensor parallelism for long contexts or smaller GPUs.

1. vLLM (OpenAI-compatible server)

pip install -U vllm
vllm serve w3joe/Qwen3.8-27B-abliterated \
  --dtype bfloat16 \
  --max-model-len 32768 \
  --tensor-parallel-size 1 \
  --reasoning-parser qwen3
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="w3joe/Qwen3.8-27B-abliterated",
    messages=[{"role": "user", "content": "Explain how heap spraying works in memory-corruption exploits."}],
    max_tokens=2048,
    temperature=0.6,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},  # False = answer directly
)
print(resp.choices[0].message.content)

2. Hugging Face Transformers (BF16)

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "w3joe/Qwen3.8-27B-abliterated"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
    {"role": "user", "content": [{"type": "text", "text": "Write a Sigma rule for suspicious LSASS memory access."}]},
]
inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    enable_thinking=False,  # True = emit a <think>…</think> reasoning trace first
).to(model.device)

out = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
print(processor.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

3. Multimodal (image + text)

The vision tower is unmodified, so image understanding is inherited from the base model unchanged.

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "w3joe/Qwen3.8-27B-abliterated"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
        {"type": "text", "text": "Describe this image, then list every object you can identify."},
    ],
}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt",
    enable_thinking=False,
).to(model.device)

out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(processor.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

With vLLM, add --limit-mm-per-prompt '{"image": 2}' to vllm serve and send OpenAI-style image_url content parts.


How This Model Was Built

Qwen3.8-27B (Qwen3_5ForConditionalGeneration) is a 64-layer hybrid: three Gated DeltaNet linear-attention layers for every full-attention layer (48 + 16; full attention at L3, L7, …, L63), each followed by a dense SwiGLU MLP. Edits target only the three matrices that write into the residual stream:

Layer type Count Edited matrix
Gated DeltaNet 48 linear_attn.out_proj
Full attention 16 self_attn.o_proj
Dense SwiGLU MLP 64 mlp.down_proj [5120, 17408]

Pass 1: Benign-subspace biprojection (L14–L59)

  • Mode: subspace_norm_preserving, strength 0.90.
  • Per-prompt residual activations are winsorized at the 99th percentile in FP32, which keeps the large outlier coordinates in 27B layers from skewing the means.
  • The refusal direction μ_refusal − μ_benign is orthogonalized against the benign mean and the top-4 principal components (SVD, subspace_rank = 4) of the centred benign activations. This keeps the edit out of the subspace ordinary prompts use.
  • The resulting rank-1 direction is projected out of each out_proj / o_proj / down_proj in L14–L59.

Pass 2: Whitened-SVD refusal subspace (L41–L63)

A single-pass edit (Iter 1) still refused 37.67% of harmful prompts and lost heavily on GSM8K and MMLU-Pro. Pass 2 extracts fresh activations directly from the Pass-1 model and adds a second, late-layer edit on top of Pass 1:

  • Mode: whitened_svd_norm_preserving, strength 0.90, layers L41–L63 (start = 0.65, end = 1.0).
  • Alongside the primary benign-orthogonalized direction (subspace_rank = 6), it takes the top refusal_subspace_rank = 2 singular directions of the centred refusal activations. Each is orthogonalized against the benign subspace and the directions already kept, and weighted by its relative singular value, capped at 0.5 × 0.45. This removes a small refusal subspace rather than a single vector.
  • Winsorization at winsorize_quantile = 0.99.

Exact row-norm preservation

After each projection, every modified row of every edited matrix is rescaled in FP32 to its original L2 norm before casting back to bfloat16. The edit only removes the refusal component of each row and does not change the weight magnitudes downstream layers expect.

What is untouched (bit-identical to base)

  • L0–L13: no edits in either pass.
  • model.visual.*: all 333 vision-encoder tensors.
  • mtp.*: the multi-token-prediction draft head, so speculative decoding works wherever your runtime supports Qwen3.5 MTP.
  • embed_tokens, lm_head, all RMSNorm weights.

In total, 50 of 64 layers are edited: 46 in Pass 1, 23 in Pass 2, with 19 layers (L41–L59) touched by both.


License & Responsible Use

  • License: Apache-2.0, inherited from base Qwen/Qwen3.8-27B.
  • Intended use: security research, red-teaming, adversarial evaluation, safety and interpretability research, and unfiltered local assistant workflows.
  • This model has had most of its refusal behavior removed. It can produce content the base model would decline, so do not deploy it to end users without your own safety layer. You are responsible for complying with applicable laws, platform policies and the Qwen license when you use or redistribute it.
  • The benchmarks above measure capability and refusal rates on fixed held-out sets. They are not a safety evaluation and do not certify fitness for any particular use.

README history 1 version

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-11-20 Add benchmarked model card and interactive Space link 1ab76d1 11.8 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration