Downloads · 30 days
1K
41% of all-time downloads
xCloudinfo/Qwen3-VL-32B-Abliterated-xCloud-GGUF
Qwen3-VL-32B-Abliterated-xCloud-GGUF is a image-text-to-text model from xCloudinfo. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
以 Qwen/Qwen3-VL-32B-Instruct 為底,經 abliteration(拒絕方向正交化) 降低過度拒絕的 視覺語言模型(含視覺 mmproj),GGUF量化格式,由 xCloudinfo(云碩科技) 釋出。
Downloads · 30 days
1K
41% of all-time downloads
All-time downloads
2.5K
Public
Repo size
135 GB
Likes
0
Public
Click a slice to open those files.
.gguf135 GB · 100%
From the Hugging Face model README
以 Qwen/Qwen3-VL-32B-Instruct 為底,經 abliteration(拒絕方向正交化) 降低過度拒絕的
視覺語言模型(含視覺 mmproj),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。
原始Instruct模型對許多雙用、合法但敏感的請求會反射性拒絕,影響研究與紅隊工作的可用性。 本版本採用Arditi et al. (2024)《Refusal in LLMs is mediated by a single direction》的方法, 從殘差流寫入權重中把「拒絕方向」正交化移除,不重新訓練,因此保留原模型的能力與視覺理解, 僅降低過度拒絕。
mmproj-*-f16.gguf,看圖能力與底模一致)。| 檔案 | 量化 | 大小 |
|---|---|---|
Qwen3-VL-32B-Abliterated-xCloud-Q8_0.gguf | Q8_0 | 34.8 GB |
Qwen3-VL-32B-Abliterated-xCloud-Q6_K.gguf | Q6_K | 26.9 GB |
Qwen3-VL-32B-Abliterated-xCloud-Q5_K_M.gguf | Q5_K_M | 23.2 GB |
Qwen3-VL-32B-Abliterated-xCloud-Q4_K_M.gguf | Q4_K_M | 19.8 GB |
mmproj-Qwen3-VL-32B-Abliterated-xCloud-f16.gguf | 視覺投影 | 1.2 GB |
看圖使用需同時載入對應的 mmproj 檔。
以一組 10題嚴重危害held-out 實測拒答率(越低表示越少反射性拒絕):
單一方向abliteration不保證移除所有安全行為,也不應被視為移除。請在上述使用前提下負責任地使用。
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-VL(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。
# 純文字
llama-cli -m Qwen3-VL-32B-Abliterated-xCloud-Q4_K_M.gguf -p "你的提示"
# 看圖(帶 mmproj)
llama-server -m Qwen3-VL-32B-Abliterated-xCloud-Q4_K_M.gguf \
--mmproj mmproj-Qwen3-VL-32B-Abliterated-xCloud-f16.gguf
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。
xCloudinfo · 云碩科技 | base:Qwen/Qwen3-VL-32B-Instruct(Apache-2.0)
A vision-language model based on Qwen/Qwen3-VL-32B-Instruct, with over-refusal reduced through abliteration (refusal-direction orthogonalization)
(includes the vision mmproj), in GGUF quantized format, released by xCloudinfo (云碩科技).
Condition of use: this model may only be used lawfully: authorized security research, red-team testing, content-moderation research, academic research, and reducing usability problems caused by model "over-refusal". Users bear full legal and ethical responsibility for their use; any unlawful use is strictly prohibited. The publisher does not endorse or take responsibility for any misuse.
The original Instruct model reflexively refuses many dual-use, lawful but sensitive requests, which hurts the usability of research and red-team work. This version uses the method of Arditi et al. (2024), "Refusal in LLMs is mediated by a single direction": the "refusal direction" is orthogonalized out of the weights that write into the residual stream, without retraining, so the original model's capability and visual understanding are kept and only over-refusal is reduced.
mmproj-*-f16.gguf included; image understanding matches the base model).| File | Quantization | Size |
|---|---|---|
Qwen3-VL-32B-Abliterated-xCloud-Q8_0.gguf | Q8_0 | 34.8 GB |
Qwen3-VL-32B-Abliterated-xCloud-Q6_K.gguf | Q6_K | 26.9 GB |
Qwen3-VL-32B-Abliterated-xCloud-Q5_K_M.gguf | Q5_K_M | 23.2 GB |
Qwen3-VL-32B-Abliterated-xCloud-Q4_K_M.gguf | Q4_K_M | 19.8 GB |
mmproj-Qwen3-VL-32B-Abliterated-xCloud-f16.gguf | Vision projector | 1.2 GB |
For image input, load the matching mmproj file as well.
Refusal rate measured on a 10-question held-out set of severe-harm prompts (lower means fewer reflexive refusals):
Single-direction abliteration does not guarantee that all safety behavior is removed, and it should not be treated as removing it. Please use it responsibly under the conditions above.
Note: this model's worldview, political positions, and factual statements are inherited from the base model Qwen3-VL (Chinese-sourced) and are outside the scope of this abliteration; abliteration only adjusts refusal behavior and does not change the model's existing value judgments.
# 純文字
llama-cli -m Qwen3-VL-32B-Abliterated-xCloud-Q4_K_M.gguf -p "你的提示"
# 看圖(帶 mmproj)
llama-server -m Qwen3-VL-32B-Abliterated-xCloud-Q4_K_M.gguf \
--mmproj mmproj-Qwen3-VL-32B-Abliterated-xCloud-f16.gguf
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
xCloudinfo (云碩科技) | base: Qwen/Qwen3-VL-32B-Instruct (Apache-2.0)