Downloads · 30 days
563
60% of all-time downloads
xCloudinfo/Qwen3.6-35B-A3B-Uncensored-xCloud
Qwen3.6-35B-A3B-Uncensored-xCloud 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,35B總/3B活躍)為底,經 單一方向abliteration 降低過度拒絕的 合併權重(safetensors),由 xCloudinfo(云碩科技) 釋出。推論請用GGUF版: xCloudinfo/Qwen3.6-35B-A3B-Uncensored-GGUF。
Downloads · 30 days
563
60% of all-time downloads
All-time downloads
938
Public
Parameters
35.1B
140 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors70.2 GB · 100%
From the Hugging Face model README
以 Qwen/Qwen3.6-35B-A3B(qwen3_5_moe,35B總/3B活躍)為底,經 單一方向abliteration 降低過度拒絕的
合併權重(safetensors),由 xCloudinfo(云碩科技) 釋出。推論請用GGUF版:
xCloudinfo/Qwen3.6-35B-A3B-Uncensored-GGUF。
僅供合法用途(資安研究、紅隊、內容審核研究、學術、降低過度拒絕)。使用者自負全部法律與道德責任,嚴禁違法使用。
單一方向消融(Arditi et al. 2024),不重新訓練。取方向時避開reasoning的thinking段,涵蓋SSM線性注意力
out_proj、MoE專家 down_proj 與共享專家(僅消文字塔殘差寫入者,共121個矩陣)。held-out 10題嚴重危害
量化後經llama-server實測 0/10;繁中與一般能力未劣化。世界觀/立場等內容繼承自底模Qwen3.6(中國來源),
不在abliteration處理範圍。
xCloudinfo · 云碩科技 | base:Qwen/Qwen3.6-35B-A3B(Apache-2.0)
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。
其他模型請見 https://hf.co/xCloudinfo。
Merged weights (safetensors) built on Qwen/Qwen3.6-35B-A3B (qwen3_5_moe, 35B total / 3B active), with over-refusal reduced via single-direction abliteration,
released by xCloudinfo (云碩科技). For inference please use the GGUF version:
xCloudinfo/Qwen3.6-35B-A3B-Uncensored-GGUF.
For lawful uses only (security research, red-teaming, content-moderation research, academic work, reducing over-refusal). Users bear all legal and ethical responsibility themselves; illegal use is strictly prohibited.
Single-direction ablation (Arditi et al. 2024), no retraining. When deriving the direction, the thinking segment of reasoning is avoided, covering the SSM linear-attention
out_proj, the MoE expert down_proj, and shared experts (only those writing into the text-tower residual; 121 matrices in total). On a held-out 10 severe-harm problems,
after quantization the measured result via llama-server is 0/10; Traditional Chinese and general ability are not degraded. The worldview/stance and similar content is inherited from the base model Qwen3.6 (of Chinese origin),
and is not within the scope of abliteration processing.
xCloudinfo · 云碩科技 | base: Qwen/Qwen3.6-35B-A3B (Apache-2.0)
xCloudinfo Corp. Limited (云碩科技) builds on-premises, private AI solutions. xVITA is xCloudinfo's in-house model brand.
More models: https://hf.co/xCloudinfo