coder3101/Qwen3.5-2B-heretic

VISIONConcurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

coder3101/Qwen3.5-2B-heretic is a 2.3 billion parameter causal language model, a decensored version of Qwen/Qwen3.5-2B created using Heretic v1.2.0 with Magnitude-Preserving Orthogonal Ablation (MPOA). This model significantly reduces refusals compared to the original, dropping from 97/100 to 7/100, while maintaining the Qwen3.5 architecture which features a unified vision-language foundation, efficient hybrid architecture, scalable RL generalization, and global linguistic coverage. It is designed for prototyping, task-specific fine-tuning, and research, particularly where reduced content moderation is desired.

Loading preview...

Model Overview

coder3101/Qwen3.5-2B-heretic is a 2.3 billion parameter multimodal causal language model, derived from Qwen/Qwen3.5-2B. Its primary distinction is being a decensored version, achieved through the Heretic v1.2.0 tool using Magnitude-Preserving Orthogonal Ablation (MPOA). This process drastically reduces model refusals from 97/100 to 7/100, as reported in the README, making it suitable for use cases requiring less content moderation.

Key Capabilities

  • Decensored Output: Significantly reduced refusal rates compared to the base Qwen3.5-2B model.
  • Multimodal Learning: Features a unified vision-language foundation, excelling in reasoning, coding, agents, and visual understanding benchmarks.
  • Efficient Architecture: Utilizes Gated Delta Networks and sparse Mixture-of-Experts for high-throughput inference.
  • Extensive Language Support: Expanded to cover 201 languages and dialects.
  • Long Context Window: Supports a native context length of 262,144 tokens.
  • Vision-Language Integration: Capable of processing image and video inputs, with strong performance across various VQA and document understanding benchmarks.

Use Cases

  • Prototyping and Research: Ideal for early-stage development and academic exploration.
  • Task-Specific Fine-tuning: Suitable as a base model for fine-tuning on specialized tasks.
  • Applications Requiring Reduced Refusals: Particularly useful in scenarios where the base model's content moderation is too restrictive.
  • Multimodal Applications: Effective for tasks involving text, image, and video understanding, such as visual question answering, document analysis, and video summarization.
  • Agentic Workflows: Excels in tool calling capabilities, recommended for use with frameworks like Qwen-Agent and Qwen Code.

Finetunes
1 models