Dingdust/Qwen3.8-27B-heretic

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 15, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Dingdust/Qwen3.8-27B-heretic is a 27 billion parameter causal language model based on the Qwen3.8 architecture, featuring native vision-language capabilities and a 32,768 token context length. This model is a decensored version, modified using Heretic v1.4.0, and is optimized for complex agentic tasks, coding, and multimodal understanding. It excels in areas like agentic terminal coding, software engineering, and long-horizon office work, offering enhanced autonomous planning and environment feedback handling.

Loading preview...

Dingdust/Qwen3.8-27B-heretic: Decensored Multimodal Agent

This model is a 27 billion parameter variant of the Qwen3.8 architecture, distinguished by its decensored nature, achieved through modification with Heretic v1.4.0. It builds upon the Qwen3.5 foundation, offering significant advancements in coding, professional tasks, research, and long-horizon agentic capabilities. The model natively supports vision-language understanding, processing both images and videos, and features flexible thinking control with adjustable reasoning depth.

Key Capabilities

  • Decensored Output: Modified to reduce refusals, with 10/100 refusals compared to 89/100 in the original model.
  • Advanced Agentic Performance: Demonstrates strong autonomous planning and improved handling of environment feedback, leading to more reliable end-to-end task completion.
  • Multimodal Understanding: Provides native support for image and video understanding, including STEM diagrams, documents, and hour-scale videos.
  • Enhanced Coding: Achieves high scores in agentic terminal coding (73.0 on Terminal Bench 2.1), agentic coding (61.7 on SWE-bench Pro, 42.2 on DeepSWE 1.1), and software engineering (79.0 on QwenSWEBench).
  • Flexible Thinking Control: Features a default 'thinking mode' that can be disabled, and reasoning depth can be tuned using reasoning_effort (xhigh, medium, low) while retaining historical reasoning context via preserve_thinking.
  • Extended Context Length: Natively supports a context length of 262,144 tokens, extensible up to 1,000,000 tokens using YaRN scaling techniques.

Good For

  • Applications requiring a decensored language model.
  • Complex agentic workflows, including coding, office work, and professional tasks.
  • Multimodal applications involving image and video analysis.
  • Scenarios demanding deep reasoning and long-horizon task completion.

Finetunes
1 models