tvall43/Qwen3.5-2B-heretic-v3b
The tvall43/Qwen3.5-2B-heretic-v3b is a 2.3 billion parameter, decensored version of the Qwen3.5-2B model, built using the Heretic v1.2.0 tool. It features a 32768 token context length and is specifically optimized to reduce refusals compared to its base model. This model is designed for use cases requiring less restrictive content generation, while maintaining the multimodal and architectural efficiencies of the original Qwen3.5 series.
Loading preview...
Model Overview
The tvall43/Qwen3.5-2B-heretic-v3b is a 2.3 billion parameter language model derived from the unsloth/Qwen3.5-2B base model. Its primary distinction is being a decensored version, achieved through the application of the Heretic v1.2.0 tool. This modification significantly reduces refusal rates, with the Heretic version showing 2 refusals out of 100 compared to 97/100 for the original model.
Key Capabilities
- Decensored Output: Engineered to produce responses with significantly fewer refusals, offering greater flexibility in content generation.
- Multimodal Foundation: Inherits the unified vision-language capabilities of the Qwen3.5 series, supporting both text and image inputs.
- Efficient Architecture: Utilizes Gated Delta Networks and sparse Mixture-of-Experts for high-throughput inference.
- Extensive Context Length: Supports a native context length of 262,144 tokens, enabling processing of very long inputs.
- Agentic Usage: Excels in tool calling, compatible with frameworks like Qwen-Agent and Qwen Code.
Good For
- Research into Content Moderation: Ideal for exploring the effects of decensoring on model behavior and output.
- Applications Requiring Unfiltered Responses: Suitable for use cases where strict content filtering is undesirable or needs to be bypassed.
- Prototyping and Development: A robust base for task-specific fine-tuning and experimental AI applications.
- Multimodal Tasks: Capable of handling both text-only and vision-language inputs, including image and video understanding.