The arnomatic/Qwen3-4B-Instruct-2507-heretic-av2 model is a 4.0 billion parameter causal language model, derived from Qwen's Qwen3-4B-Instruct-2507, and specifically modified using the Heretic v1.1.0 tool. This version is 'decensored' and 'abliterated' to significantly reduce refusals, offering enhanced capabilities in instruction following, reasoning, and long-context understanding up to 262,144 tokens. It is particularly suited for applications requiring less restrictive content generation and robust general-purpose AI tasks.
Loading preview...
arnomatic/Qwen3-4B-Instruct-2507-heretic-av2: Decensored Qwen3-4B-Instruct
This model is a specialized, "decensored" and "abliterated" variant of the Qwen3-4B-Instruct-2507 model, developed by Qwen and modified by arnomatic using the Heretic v1.1.0 tool. It retains the core capabilities of the original Qwen3-4B-Instruct-2507 while significantly reducing content refusals.
Key Enhancements & Capabilities
- Decensored Output: Achieves a refusal rate of 19/100 compared to the original's 100/100, making it suitable for use cases requiring less content filtering.
- Robust General Capabilities: Inherits and improves upon the original's strengths in instruction following, logical reasoning, text comprehension, mathematics, science, coding, and tool usage.
- Extended Context Length: Natively supports an impressive 262,144-token context window, enabling deep understanding and generation for very long inputs.
- Multilingual Proficiency: Offers substantial gains in long-tail knowledge coverage across multiple languages.
- Agentic Use: Excels in tool-calling capabilities, recommended for use with Qwen-Agent for complex task automation.
Performance Highlights
Compared to the original Qwen3-4B-Instruct-2507, this Heretic-modified version demonstrates a KL divergence of 0.0836, indicating its divergence from the original's output distribution, primarily due to the decensoring process. The base Qwen3-4B-Instruct-2507 itself shows strong performance across various benchmarks, often outperforming other models in its class, including significant gains in reasoning (e.g., AIME25, HMMT25, ZebraLogic) and alignment (e.g., Arena-Hard v2, Creative Writing v3).
When to Use This Model
This model is ideal for developers and applications that require:
- Unfiltered Content Generation: Use cases where the default content moderation of instruction-tuned models is too restrictive.
- Advanced Instruction Following: For tasks demanding precise adherence to complex instructions.
- Long-Context Understanding: Applications involving extensive documents, codebases, or conversations.
- Tool-Use and Agentic Workflows: Leveraging its strong tool-calling abilities for automated tasks.