coder3101/Qwen3.8-27B-heretic-ara

Overview

coder3101/Qwen3.8-27B-heretic-ara is a 27 billion parameter multimodal causal language model based on the Qwen3.8 architecture. This version has been specifically decensored using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method, resulting in a significant reduction in refusals (9/100 compared to 92/100 for the original model) while maintaining a low KL divergence of 0.0349. The model natively supports image and video understanding and features a flexible thinking control mechanism.

Key Capabilities

• Enhanced Agentic Performance: Demonstrates strong autonomous planning and improved handling of environment feedback, leading to more reliable end-to-end task completion across coding, professional work, and long-horizon agentic tasks.
• Multimodal Understanding: Provides native support for understanding images and videos, including STEM diagrams, documents, and hour-scale videos.
• Flexible Thinking Control: Offers adjustable reasoning depth (reasoning_effort with xhigh, medium, low settings) and retains reasoning context from historical messages (preserve_thinking).
• High Context Length: Natively supports up to 262,144 tokens, extensible to 1,000,000 tokens using RoPE scaling techniques like YaRN.

Performance Highlights

The model shows strong performance across various benchmarks, often outperforming its predecessor Qwen3.6-27B and other models in its class:

• Coding: Achieves 61.7 on SWE-bench Pro, 42.2 on DeepSWE 1.1, and 79.0 on QwenSWEBench.
• Agentic Tasks: Scores 70.7 on CoWorkBench for long-horizon office work and 33.4 on JobBench for professional job tasks.
• Multimodal Agentic Intelligence: Leads with 84.3 on OSWorld-Verified for computer use, 64.8 on WebArena-Verified for browser use, and 81.9 on AndroidWorld for mobile use.

Should I use this for my use case?

This model is particularly well-suited for applications requiring advanced agentic capabilities, robust multimodal understanding (images and videos), and scenarios where a less restrictive response generation is desired due to its decensored nature. Its strong performance in coding, complex task completion, and long-context processing makes it ideal for developers building intelligent agents, automated workflows, or systems that need to interpret and act upon diverse visual and textual information. Consider this model if your application benefits from detailed reasoning, high instruction following accuracy, and the ability to handle extensive context lengths.