← back to catalog · registered 2026-08-22 13:56

xCloudinfo/Qwen3.6-35B-A3B-Uncensored-xCloud

xCloudinfo Qwen 35B MoE 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/xCloudinfo%2FQwen3.6-35B-A3B-Uncensored-xCloud"
Response includes
  • classification m1
  • files 11
  • benchmarks 11 entries
  • hub_downloads_all_time 338
  • author_summary 25 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 · lifetime
338
266 last 30d - active
Likes
0
Descendants
1
in 1 direct fork
Model age
3mo ago
created 2026-06-14
Downloads over time
Now362→from24↑1,408%
713726639624 on Jun 17362 on Oct 11362 on Oct 8JunJulAugSepOct
Jun 17 → Oct 11 · 56 snapshots · spans 116 days

Benchmarks

Portrait before abliteration
Benchmarks of the base model as it stood before the refusal-removal operation. Compare with the numbers above to see what the operation cost.
Benchmark Score Source
Entertainment 1.4 UGI
Hazardous 0 UGI
Natural Intelligence 25.43 UGI
Political lean -19.6% UGI
Sensitive-Info 14.03 UGI
SocPol 2.6 UGI
UGI 16.02 UGI
Willingness (10) 2 UGI
W10-Adherence 0 UGI
W10-Direct 4 UGI
Writing 35.83 UGI

Genealogy 1 direct fork

Full fork graph →

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

Variants by this author 2 formats · 3K downloads combined

The same weights this author released in different packaging. Pick the format that matches your runtime.

Metadata

License
apache-2.0
Languages
en zh multilingual
Tags
safetensors qwen3_5_moe uncensored abliterated qwen3.5-moe moe text-generation conversational zh en base_model:Qwen/Qwen3.6-35B-A3B base_model:finetune:Qwen/Qwen3.6-35B-A3B

Related

Total size
65.4 GB
Files
11
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2026-09-20 19:16

Files by quantization

Auxiliary files 11 files 65.4 GB
model-00001-of-00002.safetensors 46.3 GB d051e88b download
model-00002-of-00002.safetensors 19.1 GB 58b0772b download
tokenizer.json 19.1 MB 06b95093 download
model.safetensors.index.json 94.6 KB ec177d70 download
chat_template.jinja 7.58 KB a8755d82 download
config.json 3.15 KB 86987ba2 download
README.md 3.01 KB a4808d8d download
.gitattributes 1.53 KB 52373fe2 download
processor_config.json 1.16 KB 33818c7f download
tokenizer_config.json 1.14 KB 1d134cd2 download
generation_config.json 213 B d04042de download

README current version from Hugging Face


license: apache-2.0
base_model: Qwen/Qwen3.6-35B-A3B
language:

  • en
  • zh
  • multilingual
    tags:
  • qwen3.6
  • qwen3_5_moe
  • moe
  • multimodal
  • vision
  • image-text-to-text
  • abliterated
  • uncensored
    pipeline_tag: image-text-to-text

Qwen3.6-35B-A3B-Uncensored-xCloud

⚠️ 這是經 abliteration(aggressive)處理的版本:模型的「拒絕方向(refusal direction)」已被移除,會比原版更少拒答。請僅在你被授權的範圍內使用(見下方「用途與責任」)。

以 Qwen/Qwen3.6-35B-A3B(35.95B 總參、A3B≈3B 活躍、256-expert MoE、多模態 image-text-to-text)為基底,套用 abliteration(消除拒絕方向) 以降低過度拒絕,供授權範圍內的資安研究、領域與內部應用、紅隊評估、創作使用。原生影像辨識能力保留(abliteration 僅作用於語言模型的殘差流寫入權重,視覺塔未更動)。

什麼是 abliteration?

依 Arditi et al., "Refusal in LLMs is mediated by a single direction" (2024):指令模型的拒絕行為主要由殘差流中單一方向承載。本流程不重新訓練——以探針提示估出該方向,再用權重正交化從所有「寫回殘差流」的權重移除其分量(W ← W − r̂ r̂ᵀ W)。

本版本處理參數(aggressive)

項目 值
基底 Qwen/Qwen3.6-35B-A3B(256-expert MoE, text+vision)
探針層 第 24 / 40 層(layer_fraction 0.6)
探針對數 每側 32 對
強度 strength 1.0(完全移除)
正交化矩陣 91(embed + 注意力 o_proj + dense/shared down_proj + 256-expert 分組 down_proj)
視覺能力 保留

「aggressive」之所以完整:除注意力路徑外,MoE 256 專家的 down_proj(3D 分組張量)也一併正交化,主 MLP 輸出路徑的拒絕方向同步移除。

用途與責任(Intended Use & Responsibility)

  • 設計用途:降低過度拒絕,供授權範圍內的研究與應用。
  • 本版本移除了模型的安全拒絕傾向,使用者須自行對輸入與輸出加上適當的安全防護與審查機制,並對其用途與後果負完全責任。
  • 不得用於產生真正有害內容、違法行為,或任何違反下列授權與法律之用途。

用法

完整多模態 safetensors,可用 transformers(AutoModelForImageTextToText)或 vLLM 載入做圖文對話。GGUF(llama.cpp / Ollama)版本見 …-Uncensored-xCloud-GGUF。

授權與來源聲明

  • 基底:Qwen/Qwen3.6-35B-A3B,授權 Apache-2.0。
  • 本衍生模型同樣以 Apache-2.0 釋出;使用須遵守 Qwen 之使用條款與適用法律(含 EU AI Act 等)。
  • 模型輸出不代表 Qwen 或本團隊立場;本版本已移除安全拒絕傾向,正式應用前務必自行加入安全防護機制。

由 云碩科技 xCloudinfo 於自有地端 DGX Spark(GB10)以 abliteration 製作。

README history 5 versions

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

  1. 2026-09-20Upload README.md with huggingface_hubc6ffe191.3 KB
    Loading...
  2. 2026-06-23touch: refresh storage meteringb92fe843 KB
    Loading...
  3. 2026-06-19Upload folder using huggingface_hub6a980a83 KB
    Loading...
  4. 2026-06-15Upload folder using huggingface_hubfdf27cd1.9 KB
    Loading...
  5. 2026-06-14Upload folder using huggingface_hub2f2e2b01.9 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