ArliAI/Qwen3.5-27B-Derestricted

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Mar 8, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

ArliAI/Qwen3.5-27B-Derestricted is a 27 billion parameter multimodal causal language model developed by Arli AI, based on Qwen3.5-27B. This model has been specifically modified using Norm-Preserving Biprojected Abliteration to remove refusal behaviors while preserving and potentially improving the original model's reasoning capabilities. It excels in multimodal understanding, supporting text, image, and video inputs, and features an extended context length of 32,768 tokens, extensible up to 1,010,000 tokens with YaRN scaling.

Loading preview...

Overview

ArliAI/Qwen3.5-27B-Derestricted is a 27 billion parameter multimodal causal language model developed by Arli AI. It is a "derestricted" version of the original Qwen3.5-27B, engineered to remove refusal behaviors without degrading performance. This is achieved through a novel technique called Norm-Preserving Biprojected Abliteration, which carefully modifies the model's weights to remove refusal components while preserving the neural network's delicate feature norms and potentially enhancing reasoning.

Key Capabilities

  • Refusal-Free Operation: Designed to eliminate refusal behaviors commonly found in base models, offering more direct and unfiltered responses.
  • Multimodal Understanding: Supports unified processing of text, image, and video inputs, demonstrating strong performance across various vision-language benchmarks.
  • Extended Context Length: Natively handles up to 32,768 tokens, with extensibility to 1,010,000 tokens using YaRN scaling for ultra-long text processing.
  • Enhanced Reasoning: Benchmarks suggest that by removing the "safety tax" of refusal behaviors, the model may exhibit improved reasoning capabilities compared to its baseline.
  • Agentic Functionality: Excels in tool calling, with recommended integration via Qwen-Agent and Qwen Code for building sophisticated AI applications.

Good for

  • Applications requiring unfiltered and direct responses from an LLM.
  • Multimodal tasks involving complex reasoning over text, images, and videos.
  • Developing AI agents that leverage tool-use capabilities.
  • Scenarios demanding long-context understanding, such as document analysis or extended conversations.

Finetunes
2 models