richardyoung/Qwen2.5-3B-Instruct-heretic

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 24, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The richardyoung/Qwen2.5-3B-Instruct-heretic is a 3.09 billion parameter instruction-tuned causal language model, based on the Qwen2.5 architecture, with a 32,768 token context length. This model is a decensored version of Qwen/Qwen2.5-3B-Instruct, created using Heretic v1.4.0, significantly reducing refusals compared to the original. It retains the Qwen2.5 improvements in coding, mathematics, instruction following, and structured output generation, making it suitable for applications requiring less content moderation.

Loading preview...

richardyoung/Qwen2.5-3B-Instruct-heretic: Decensored Qwen2.5-3B-Instruct

This model is a 3.09 billion parameter instruction-tuned causal language model, derived from the Qwen2.5 architecture, and offers a 32,768 token context length. It is a decensored variant of the original Qwen/Qwen2.5-3B-Instruct, created using Heretic v1.4.0.

Key Differentiators & Capabilities

  • Decensored Output: Significantly reduces refusals, with only 3 out of 100 refusals compared to 96 out of 100 in the original model, as measured by KL divergence of 0.0494.
  • Enhanced Knowledge & Reasoning: Inherits Qwen2.5's improvements in coding and mathematics, leveraging specialized expert models.
  • Instruction Following: Demonstrates significant improvements in instruction following and generating long texts (over 8K tokens).
  • Structured Data & Output: Excels at understanding structured data (e.g., tables) and generating structured outputs, particularly JSON.
  • Robust Prompt Handling: More resilient to diverse system prompts, enhancing role-play and condition-setting for chatbots.
  • Multilingual Support: Supports over 29 languages, including Chinese, English, French, Spanish, Portuguese, German, Italian, Russian, Japanese, and Korean.

Use Cases

This model is particularly well-suited for applications where reduced content moderation and direct responses are preferred, while still benefiting from the strong capabilities of the Qwen2.5 base model in coding, mathematics, and instruction following.

Finetunes
1 models