richardyoung/Qwen2.5-Coder-14B-Instruct-heretic

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 24, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Warm

The richardyoung/Qwen2.5-Coder-14B-Instruct-heretic is a 14.8 billion parameter instruction-tuned causal language model, derived from the Qwen2.5-Coder-14B-Instruct by Qwen, and modified using the Heretic v1.4.0 tool. This model specializes in code generation, reasoning, and fixing, building upon a training dataset of 5.5 trillion tokens including extensive source code. It offers a 32K context length, with support for up to 128K tokens via YaRN, making it suitable for complex coding tasks and long-context applications.

Loading preview...

Model Overview

This model, richardyoung/Qwen2.5-Coder-14B-Instruct-heretic, is a 14.8 billion parameter instruction-tuned causal language model. It is a modified version of the original Qwen/Qwen2.5-Coder-14B-Instruct from Qwen, processed with the Heretic v1.4.0 tool to be a "decensored" variant. The base Qwen2.5-Coder series is designed for code-specific tasks, leveraging a massive 5.5 trillion token training corpus that includes source code and text-code grounding.

Key Capabilities

  • Enhanced Code Performance: Significantly improved capabilities in code generation, code reasoning, and code fixing compared to its predecessors.
  • Long Context Support: Features a native context length of 32,768 tokens, with support for up to 131,072 tokens using YaRN (Yet another RoPE Normalization) for handling extensive inputs.
  • General Competencies: While specialized for coding, it maintains strong performance in mathematics and general language understanding.
  • Reproducibility: The modification process using Heretic v1.4.0 is stated to be reproducible.

Differentiator

This "heretic" version specifically aims to reduce refusals, showing a significant drop from 100/100 in the original model to 3/100 in this variant, as measured by KL divergence and refusal metrics. This makes it distinct for use cases requiring less restrictive content generation.

Use Cases

This model is well-suited for advanced code-related applications, including code generation, debugging, and acting as a Code Agent, especially in scenarios where a less restrictive output policy is desired.

Finetunes
1 models