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

Model Overview

This model, heretic-org/Qwen3.8-27B-heretic-ara, is a 27 billion parameter causal language model with a vision encoder, built upon the Qwen3.8 architecture. It has been decensored using the Heretic tool with the Arbitrary-Rank Ablation (ARA) method, specifically targeting layers 26 through 56 to reduce refusals from 99/100 in the original model to 0/100 in this version.

Key Capabilities & Enhancements

• Decensored Behavior: Significantly reduced refusal rates compared to the base Qwen3.8-27B model, achieved through targeted ablation.
• Multimodal Understanding: Native support for interpreting images and videos, including STEM diagrams, documents, and hour-long video content.
• Advanced Agent Execution: Demonstrates strong autonomous planning and improved handling of environmental feedback, leading to more reliable completion of complex, multi-step tasks.
• Flexible Thinking Control: Features a configurable 'thinking mode' with adjustable reasoning depth (reasoning_effort levels: xhigh, medium, low) and preserved reasoning context (preserve_thinking).
• Extended Context Length: Natively supports up to 262,144 tokens, extensible to 1,000,000 tokens using RoPE scaling techniques like YaRN.

Performance Highlights

This model shows strong performance across various benchmarks, often outperforming its base model and other competitors in its class:

• Coding: Achieves 61.7 on SWE-bench Pro, 42.2 on DeepSWE 1.1, and 79.0 on QwenSWEBench, indicating robust agentic coding capabilities.
• Agentic Tasks: Scores 70.7 on CoWorkBench for long-horizon office work and 33.4 on JobBench for professional tasks, with a Pass@1 of 20.4 and Score of 42.9 on Agents' Last Exam.
• Multimodal Agentic Intelligence: Leads with 84.3 on OSWorld-Verified (computer use), 64.8 on WebArena-Verified (browser use), and 81.9 on AndroidWorld (mobile use).
• General Multimodal Intelligence: Achieves 94.6 (with CI) on MathVision for visual math problem-solving and 85.6 (with CI) on BabyVision for general visual reasoning.

When to Use This Model

This model is particularly well-suited for applications requiring:

• Unfiltered Responses: Ideal for use cases where the original model's refusal behavior is undesirable.
• Complex Agentic Workflows: Its enhanced planning and execution capabilities make it suitable for automating multi-step tasks in coding, office work, and professional domains.
• Multimodal Data Processing: Excellent for tasks involving the understanding and generation of content from both text and visual inputs (images and videos).
• Long Context Processing: Capable of handling very long inputs, making it valuable for detailed analysis and summarization of extensive documents or video transcripts.