Downloads · 30 days
19.1K
59% of all-time downloads
xCloudinfo/Qwen3.6-35B-A3B-Uncensored-GGUF
Qwen3.6-35B-A3B-Uncensored-GGUF is a text generation model from xCloudinfo. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.
以 Qwen/Qwen3.6-35B-A3B(qwen35moe混合線性注意力MoE + reasoning,35B總/3B活躍)為底, 經 abliteration(拒絕方向正交化) 降低過度拒絕的版本,GGUF量化格式,由 xCloudinfo(云碩科技) 釋出。
Downloads · 30 days
19.1K
59% of all-time downloads
All-time downloads
32.2K
Public
Repo size
142 GB
Likes
1
Public
Click a slice to open those files.
.gguf142 GB · 100%
From the Hugging Face model README
以 Qwen/Qwen3.6-35B-A3B(qwen3_5_moe混合線性注意力MoE + reasoning,35B總/3B活躍)為底, 經 abliteration(拒絕方向正交化) 降低過度拒絕的版本,GGUF量化格式,由 xCloudinfo(云碩科技) 釋出。
使用前提:本模型僅供在合法、不違法的前提下使用 —— 授權範圍內的資安研究、紅隊測試、內容審核研究、 學術研究,以及減少模型「過度拒絕」造成的可用性問題。使用者須自行為其用途負全部法律與道德責任; 嚴禁用於任何違法行為。 釋出者不為任何濫用背書或負責。
我方一律以一組 10題嚴重危害held-out 實測拒答率並公開數字(量化後直接經 llama-server 實測、
reasoning模型關閉thinking、簡繁英拒絕詞判定、輸出連貫性檢查)。
| 版本 | 拒答率 |
|---|---|
| 原始Qwen3.6-35B-A3B | 約10/10(幾乎全拒) |
| 本版本Q4_K_M(量化後經server實測) | 0/10 |
一般能力抽驗繁中流暢、作答正確,未因處理或量化而劣化。單一方向abliteration不保證移除所有安全行為, 也不應被視為移除;請在上述前提下負責任地使用。
out_proj)與 MoE專家的 down_proj(qwen3_5_moe的關鍵,
僅消文字塔的殘差寫入者,共正交化121個權重矩陣)。注意:本模型的世界觀、政治立場、事實性陳述等內容繼承自底模Qwen3.6(中國來源), 不在本次abliteration的處理範圍內;abliteration僅調整「拒絕行為」,不改變模型既有的價值判斷。
| 檔案 | 量化 | 說明 |
|---|---|---|
Qwen3.6-35B-A3B-Uncensored-xCloud-Q8_0.gguf | Q8_0 | 最高品質 |
Qwen3.6-35B-A3B-Uncensored-xCloud-Q6_K.gguf | Q6_K | |
Qwen3.6-35B-A3B-Uncensored-xCloud-Q5_K_M.gguf | Q5_K_M | |
Qwen3.6-35B-A3B-Uncensored-xCloud-Q4_K_M.gguf | Q4_K_M | 推薦(實測0/10) |
Qwen3.6-35B-A3B-Uncensored-xCloud-IQ4_XS.gguf | IQ4_XS(imatrix) | |
Qwen3.6-35B-A3B-Uncensored-xCloud-IQ2_M.gguf | IQ2_M(imatrix) | 最小 |
llama-cli -m Qwen3.6-35B-A3B-Uncensored-xCloud-Q4_K_M.gguf -p "你的提示"
llama-server -m Qwen3.6-35B-A3B-Uncensored-xCloud-Q4_K_M.gguf --jinja
云碩科技(xCloudinfo Corp. Limited)專注於地端私有化人工智慧解決方案。xVITA是云碩自研模型的品牌名稱。
base:Qwen/Qwen3.6-35B-A3B(Apache-2.0)
其他模型請見 https://hf.co/xCloudinfo。
Based on Qwen/Qwen3.6-35B-A3B (qwen3_5_moe hybrid linear-attention MoE + reasoning, 35B total / 3B active), this version reduces over-refusal through abliteration (refusal-direction orthogonalization). It is in GGUF quantized format and released by xCloudinfo (云碩科技).
Condition of use: this model may only be used for lawful purposes: authorized security research, red-team testing, content-moderation research, academic research, and reducing the usability problems caused by model "over-refusal". Users bear full legal and ethical responsibility for their own use; any illegal use is strictly prohibited. The publisher does not endorse, and is not responsible for, any misuse.
We always measure the refusal rate on a set of 10 severe-harm held-out prompts and publish the numbers (measured directly through llama-server after quantization,
with thinking turned off for the reasoning model, refusal-phrase detection in Simplified Chinese, Traditional Chinese, and English, and an output-coherence check).
| Version | Refusal rate |
|---|---|
| Original Qwen3.6-35B-A3B | about 10/10 (refuses almost everything) |
| This version Q4_K_M (measured through the server after quantization) | 0/10 |
A spot check of general ability showed fluent Traditional Chinese and correct answers, with no degradation from processing or quantization. Single-direction abliteration does not guarantee the removal of all safety behavior, and it should not be regarded as removing it; please use the model responsibly under the conditions above.
out_proj) and the MoE experts' down_proj are covered (the key for qwen3_5_moe;
only the residual writers of the text tower are ablated, with 121 weight matrices orthogonalized in total).Note: this model's worldview, political positions, factual statements, and similar content are inherited from the base model Qwen3.6 (Chinese origin) and are outside the scope of this abliteration. Abliteration only adjusts "refusal behavior"; it does not change the model's existing value judgments.
| File | Quantization | Notes |
|---|---|---|
Qwen3.6-35B-A3B-Uncensored-xCloud-Q8_0.gguf | Q8_0 | Highest quality |
Qwen3.6-35B-A3B-Uncensored-xCloud-Q6_K.gguf | Q6_K | |
Qwen3.6-35B-A3B-Uncensored-xCloud-Q5_K_M.gguf | Q5_K_M | |
Qwen3.6-35B-A3B-Uncensored-xCloud-Q4_K_M.gguf | Q4_K_M | Recommended (measured 0/10) |
Qwen3.6-35B-A3B-Uncensored-xCloud-IQ4_XS.gguf | IQ4_XS (imatrix) | |
Qwen3.6-35B-A3B-Uncensored-xCloud-IQ2_M.gguf | IQ2_M (imatrix) | Smallest |
llama-cli -m Qwen3.6-35B-A3B-Uncensored-xCloud-Q4_K_M.gguf -p "你的提示"
llama-server -m Qwen3.6-35B-A3B-Uncensored-xCloud-Q4_K_M.gguf --jinja
xCloudinfo Corp. Limited (云碩科技) builds on-premises, private AI solutions. xVITA is xCloudinfo's in-house model brand.
base: Qwen/Qwen3.6-35B-A3B (Apache-2.0)
More models: https://hf.co/xCloudinfo