jakejb53/Qwen3.8-27B-heretic
jakejb53/Qwen3.8-27B-heretic is a 27 billion parameter, decensored version of the Qwen3.8-27B model, built using Heretic v1.4.0. This causal language model with a vision encoder offers comprehensive improvements across coding, professional work, research, and long-horizon agentic tasks, with a native context length of 262,144 tokens. Its primary differentiator is its decensored nature, alongside enhanced agent execution, flexible thinking control, and native support for image and video understanding.
Loading preview...
Model Overview
jakejb53/Qwen3.8-27B-heretic is a 27 billion parameter, decensored variant of the Qwen3.8-27B model, created using Heretic v1.4.0. This model is a causal language model with a native vision encoder, offering a substantial context length of 262,144 tokens, extensible up to 1,000,000 tokens with YaRN scaling. It is designed for advanced capabilities across various domains, including coding, professional tasks, research, and complex agentic workflows.
Key Differentiators
- Decensored Performance: Compared to the original Qwen3.8-27B, this Heretic-modified version shows 0 refusals out of 100 test cases, significantly lower than the original's 63 refusals, while maintaining a low KL divergence of 0.0236.
- Enhanced Agentic Capabilities: The model demonstrates stronger autonomous planning and improved handling of environmental feedback, leading to more reliable end-to-end task completion. It excels in benchmarks like SWE-bench Pro (61.7), DeepSWE 1.1 (42.2), QwenSWEBench (79.0), CoWorkBench (70.7), and JobBench (33.4).
- Multimodal Understanding: It natively supports image and video understanding, capable of processing STEM diagrams, documents, and hour-scale videos. Benchmarks show strong performance in agentic multimodal intelligence tasks such as OSWorld-Verified (84.3), WebArena-Verified (64.8), and AndroidWorld (81.9).
- Flexible Thinking Control: Qwen3.8-27B operates in a "thinking mode" by default, generating internal reasoning before the final response. Users can adjust reasoning depth with
reasoning_effort(xhigh, medium, low) and preserve thinking context across turns withpreserve_thinking.
Use Cases
- Complex Coding Tasks: Particularly strong in agentic coding and software engineering, outperforming several comparable models in benchmarks like SWE-bench Pro and QwenSWEBench.
- Long-Horizon Agentic Workflows: Ideal for tasks requiring autonomous planning, environment interaction, and multi-step completion in professional and research settings.
- Multimodal Applications: Suitable for applications requiring understanding and processing of both text and visual inputs, including images and videos, for tasks like visual reasoning, document intelligence, and real-world perception.