hfmaster/Qwen3-VL-2B

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Nov 14, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The hfmaster/Qwen3-VL-2B is a 2 billion parameter vision-language model, developed by huihui-ai, that has been abliterated to remove safety filtering from its text generation capabilities. Based on the Qwen3-VL-2B-Instruct architecture, this model is designed for multimodal tasks, specifically image description and analysis, without the refusal behaviors typically found in instruction-tuned models. It supports a context length of 32768 tokens and is optimized for research and experimental use where unfiltered text outputs are desired.

Loading preview...

Overview

This model, hfmaster/Qwen3-VL-2B, is an abliterated version of the Qwen/Qwen3-VL-2B-Instruct, developed by huihui-ai. It features 2 billion parameters and a context length of 32768 tokens. The abliteration process specifically targeted the text generation component, removing safety filtering and refusal behaviors, meaning the model will no longer decline to describe or analyze images based on content restrictions. The image processing capabilities remain unchanged from the original Qwen3-VL-2B-Instruct.

Key Capabilities

  • Uncensored Text Generation: Provides unfiltered descriptions and analyses of images, bypassing typical safety mechanisms.
  • Multimodal Understanding: Capable of processing both image and text inputs for comprehensive visual language tasks.
  • Ollama and GGUF Support: Easily deployable with Ollama (v0.12.7+) and compatible with GGUF format for use with llama.cpp tools like llama-mtmd-cli and llama-cli.
  • Python Integration: Demonstrates straightforward integration using the transformers library for image-to-text generation.

Use Cases and Warnings

This model is primarily intended for research and experimental use in controlled environments. Due to the removal of safety filtering, it carries a risk of generating sensitive, controversial, or inappropriate content. Users are strongly advised to exercise caution, rigorously review outputs, and understand that the model is not suitable for public-facing commercial applications or audiences requiring high security. Users bear full responsibility for ensuring compliance with legal and ethical standards.

Finetunes
1 models