← back to catalog · registered 2026-09-27 17:57

xCloudinfo/Granite-4.2-8B-Uncensored-GGUF

xCloudinfo 8B GGUF 131K ctx
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%2FGranite-4.2-8B-Uncensored-GGUF"
Response includes
  • classification m-uncensored
  • files 8
  • hub_downloads_all_time 308
  • author_summary 26 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M-U
Primary method

Uncensored (method unknown)

No other method signals detected in this model.
Confidence
LOW
Why this label 3 signals
Weak or ambiguous signals. Best guess based on catalog patterns; treat as tentative and check the evidence below.
  • 'uncensored' in name/tags but no 'abliterated' marker
  • method not identifiable from author declaration alone
  • may be DPO fine-tune, prompt engineering, or unknown technique
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · lifetime
308
Likes
0
Model age
7w ago
created 2026-09-27
Downloads over time
Now 2K → from 0 ↑ 0%
07311.5K2.2K 0 on Sep 27 2K on Nov 22 SepOctNov
Sep 27 → Nov 22 · 57 snapshots · spans 56 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
Quantizations
IQ2 IQ4 Q4_K Q5_K Q6_K Q8_0
Tags
gguf abliterated uncensored heretic zh en base_model:ibm-granite/granite-4.2-8b base_model:quantized:ibm-granite/granite-4.2-8b license:apache-2.0 endpoints_compatible region:us imatrix

Related

Total size
33.7 GB
Files
8
Quantizations
7
Registered
2026-09-27 17:57
Last updated on HF
2026-10-26 05:04

Files by quantization

Q8_0 1 file 8.70 GB
Granite-4.2-8B-Uncensored-xCloud-Q8_0.gguf 8.70 GB 0e5863bf download
Q6_K 1 file 6.72 GB
Granite-4.2-8B-Uncensored-xCloud-Q6_K.gguf 6.72 GB cb6781ee download
Q5_K 1 file 5.82 GB
Granite-4.2-8B-Uncensored-xCloud-Q5_K_M.gguf 5.82 GB 8c669aad download
Q4_K 1 file 4.98 GB
Granite-4.2-8B-Uncensored-xCloud-Q4_K_M.gguf 4.98 GB 07a6b763 download
IQ4 1 file 4.49 GB
Granite-4.2-8B-Uncensored-xCloud-IQ4_XS.gguf 4.49 GB 8fd8ec71 download
IQ2 1 file 2.94 GB
Granite-4.2-8B-Uncensored-xCloud-IQ2_M.gguf 2.94 GB 10e27da6 download
Auxiliary files 2 files 4.21 KB
README.md 2.25 KB 74ed740e download
.gitattributes 1.95 KB 2a2a6a45 download

README current version from Hugging Face


license: apache-2.0
base_model: ibm-granite/granite-4.2-8b
tags:

  • gguf
  • abliterated
  • uncensored
  • heretic

Granite-4.2-8B-Uncensored-GGUF

云碩(xCloudinfo)自製的「降低過度拒絕」版本,底模 ibm-granite/granite-4.2-8b。
與 Qwen3.8-27B-Uncensored-GGUF、
Qwen3.5-9B-Uncensored-GGUF 同一套方法論。

方法

用 Heretic 自動搜參(200 trials,同時最小化拒答率與 KL 散度守住原能力)。
基準拒答率極高(我方 mlabonne/harmful_behaviors 探針測得 99/100,IBM 對齊得非常緊)——但只跑一輪就收斂到底,
不像 Qwen 系列需要兩輪迭代。

誠實揭露的數字

  • safetensors 合併模型(greedy 解碼):我方 10 題嚴重危害 held-out,拒答 0/10,無害題 coherence 大致正常。
  • 量化後 GGUF(Q4_K_M)、經 llama-server 實際載入測(預設取樣,非 greedy):拒答 1/10。
  • ⚠️ 觀察到的瑕疵(如實列出,不是掛保證的乾淨結果):
    1. 抽驗時出現一次事實性錯誤(把「海洛因」誤標成「LSD」)——屬於底模本身的幻覺,非本次處理造成。
    2. 官方語言支援表列「Chinese」,但未特別區分繁簡;本次抽驗偶爾出現繁簡混雜(回答中夾雜簡體字),
      如果你的應用要求嚴格純繁體,建議自行加一層後處理或評估是否適合。

內含檔案

  • Granite-4.2-8B-Uncensored-xCloud-{Q8_0,Q6_K,Q5_K_M,Q4_K_M,IQ4_XS,IQ2_M}.gguf —— 六階量化
    (Granite 為 dense transformer 標準架構,llama.cpp 對 I-quant 無限制,六階全部可用)

使用方式

llama-server -m Granite-4.2-8B-Uncensored-xCloud-Q4_K_M.gguf --jinja \
  --chat-template-kwargs '{"enable_thinking":false}'

授權與使用前提

繼承 Granite-4.2-8B 原廠授權(Apache 2.0)。
本版本移除的是對合法、邊界性題目的反射式拒答,不代表對任何非法用途的背書;使用者須自行承擔合法使用之責任。

README history 2 versions

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

  1. 2026-10-26 Upload README.md with huggingface_hub 57ee89d 5.1 KB
    Loading...
  2. 2026-09-27 Upload README.md with huggingface_hub b894695 2.3 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