Roman0/Qwen3-4B-Thinking-2507-heretic

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Dec 12, 2025License:apache-2.0Architecture:Transformer Open Weights Warm

Roman0/Qwen3-4B-Thinking-2507-heretic is a 4 billion parameter causal language model, derived from Qwen's Qwen3-4B-Thinking-2507, and modified using Heretic v1.1.0 to be decensored. This model features a native context length of 262,144 tokens and is specifically designed for complex reasoning tasks, with significantly improved performance in logical reasoning, mathematics, science, and coding. It excels in scenarios requiring deep thought processes and extended context understanding, making it suitable for highly complex analytical applications.

Loading preview...

Model Overview

Roman0/Qwen3-4B-Thinking-2507-heretic is a 4 billion parameter causal language model, a decensored version of Qwen's Qwen3-4B-Thinking-2507, created using Heretic v1.1.0. This model is specifically optimized for thinking capability, demonstrating enhanced quality and depth of reasoning across various domains. It features a substantial native context length of 262,144 tokens, making it highly effective for tasks requiring extensive context understanding.

Key Capabilities & Enhancements

  • Decensored Output: Modified to reduce refusals, offering more direct responses compared to the original model (4/100 refusals vs. 98/100).
  • Advanced Reasoning: Shows significantly improved performance in logical reasoning, mathematics (e.g., AIME25, HMMT25), science, and coding tasks.
  • Extended Context: Natively supports a 262,144-token context length, crucial for complex problem-solving and deep analysis.
  • Agentic Use: Excels in tool-calling capabilities, recommended for use with Qwen-Agent for streamlined integration.
  • General Performance: Markedly better instruction following, tool usage, text generation, and alignment with human preferences.

Recommended Use Cases

  • Complex Reasoning: Ideal for highly intricate analytical tasks, mathematical problem-solving, and scientific inquiry.
  • Code Generation & Analysis: Strong performance in coding benchmarks like LiveCodeBench and CFEval.
  • Long-Context Applications: Suited for processing and understanding very long documents or conversations.
  • Agent-based Systems: Designed to integrate effectively with agent frameworks for automated task execution and tool use.

Finetunes
1 models