Abliterated text path
According to the model card, this is an uncensored version of Qwen/Qwen3-VL-4B-Instruct created with abliteration, applied to the text path rather than the image path.
Open Source Model Profile · huihui-ai
Huihui-Qwen3-VL-4B-Instruct-abliterated
Huihui-Qwen3-VL-4B-Instruct-abliterated is a 4.44B-parameter Qwen3-VL fine-tune from huihui-ai. According to the model card, it is an uncensored abliteration of Qwen3-VL-4B-Instruct with only the text path processed.
- Publisher
- huihui-ai
- Task
- image-text-to-text
- Model type
- qwen3_vl
- License
- apache-2.0
- Library
- transformers
- Publication status
- Accepted · not indexed
Model overview
Huihui-Qwen3-VL-4B-Instruct-abliterated is published by huihui-ai as an image-text-to-text model. The captured configuration identifies Qwen3VLForConditionalGeneration with model type qwen3_vl, and Safetensors metadata reports 4,437,815,808 parameters. According to the model card, it is an uncensored abliterated version of Qwen/Qwen3-VL-4B-Instruct with only the text path processed.
Recorded capabilities
Removed image refusal
According to the model card, the abliterated model no longer refuses with the original image-analysis refusal message.
Captured 4.44B Transformers weights
Safetensors metadata reports 4,437,815,808 parameters, and the hub record lists Transformers support.
Ollama and Transformers paths
According to the model card, the release is runnable through an Ollama tag and through Qwen3VLForConditionalGeneration with an AutoProcessor chat-template workflow.
Explicit safety warnings
According to the model card, safety filtering is significantly reduced with possible sensitive or controversial output, and users bear legal and ethical responsibility for review and compliance.
Use cases in the source record
- Vision-language generation experiments where the publisher describes removal of the original image-analysis refusal.
- Chat-with-image workflows through the publisher-documented Ollama tag or Transformers chat-template inference path.
Limitations and unknowns
- Abliteration effects are publisher-described changes and were not independently verified by Ethen.
- According to the model card, reduced safety filtering can produce sensitive or inappropriate content requiring rigorous user review.
- No evaluation results were extracted from this record.
- No context-window value was extracted from this record.
- Provider state is historical snapshot data and should be refreshed before being presented as current.
Source and provenance
Source: huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated
Captured: Unknown. Processed: 2026-09-07T19:34:46.669900+00:00.
huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated This is an uncensored version of Qwen/Qwen3-VL-4B-Instruct created with abliteration (see remove-refusals-with-transformers to know more about it). It was only the text part that was processed, not the image part. The abliterated model will no longer say "I can’t describe or analyze this image." ollama Please update to the latest version of Ollama-v0.12.7 . You can use huihui_ai/qwen3-vl-abliterated:4b-instruct directly, ollama run huihui_ai/qwen3-vl-abliterated:4b-instruct Chat with Image from transformers import Qwen3VLForConditionalGeneration, AutoProcessor, BitsAndBytesConfig imp…
F001F002F003F004F005F006F007F008F009F010F011F012F013F014F015F016F017