kang926/Qwopus3.6-27B-v2-heretic

Hugging Face
VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 29, 2026Architecture:Transformer0.0K Warm

kang926/Qwopus3.6-27B-v2-heretic is a 27 billion parameter language model based on the Jackrong/Qwopus3.6-27B-v2 architecture, processed using the p-e-w/heretic tool. This model is a full merged safetensors export, specifically optimized for instruction following and long-context recall, with a context length of 32768 tokens. It demonstrates capabilities in Korean instruction following, JSON-format adherence, and consistent sampling behavior, making it suitable for applications requiring reliable output and extended memory.

Loading preview...

Model Overview

The kang926/Qwopus3.6-27B-v2-heretic is a 27 billion parameter language model derived from the Jackrong/Qwopus3.6-27B-v2 base model. It has been processed and exported as a full merged safetensors file using the p-e-w/heretic tool, specifically Heretic version 1.3.0, with no quantization applied during the process. The model was developed with a focus on refining its behavior, as indicated by a low refusal rate of 5/100 and a KL divergence of 0.0070 during its build.

Key Capabilities

  • Instruction Following: The model has been manually checked for its ability to follow instructions, including specific Korean instructions.
  • Long-Context Recall: It demonstrates proficiency in recalling information over extended contexts, supporting a context length of 32768 tokens.
  • Structured Output: The model is capable of adhering to specific output formats, such as JSON.
  • Sampling Consistency: It exhibits consistent behavior across repeated sampling attempts.

Good For

  • Applications requiring robust instruction following, particularly in Korean.
  • Tasks that benefit from long-context understanding and information retrieval.
  • Use cases demanding reliable and consistent structured output, like JSON generation.
  • Scenarios where avoiding unnecessary clarification and long-answer repetition is crucial.