leimroth-lab/Qwen3.8-27B-Leimroth3

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Qwen3.8-27B-Leimroth3 is a 27 billion parameter multimodal language model derived from Qwen/Qwen3.8-27B by Scott Leimroth. This model has undergone an 'abliteration' process to significantly reduce refusal behavior by inverting the refusal direction in its language layers, while preserving its vision capabilities. It is intended for research into alignment, refusal mechanisms, and red-teaming, offering a model that declines far less often than its base counterpart.

Loading preview...

Overview

leimroth-lab/Qwen3.8-27B-Leimroth3 is a 27 billion parameter multimodal model, a modified version of Qwen/Qwen3.8-27B by Scott Leimroth. Its primary distinction is the application of an "abliteration" technique, where a single refusal direction is inverted (at lambda 1.5) within the language-layer writer matrices. This modification aims to make the model decline prompts far less often, while its vision tower remains untouched.

Key Capabilities & Characteristics

  • Reduced Refusal Behavior: The model is engineered to significantly reduce explicit refusals, making it more compliant with a wider range of prompts. Evaluation shows 90.6% plain compliance on a held-out harmful set, though this is an upper-bound due to short continuation lengths.
  • Multimodal: Retains the vision capabilities of the base Qwen model, allowing for image understanding. However, it requires disabling a "thinking step" for reliable image responses.
  • Research Focus: Specifically released for research into AI alignment, refusal mechanisms, and red-teaming, providing a tool to explore the boundaries of model compliance.
  • Performance Considerations: While designed for compliance, the abliteration process introduces some trade-offs. It runs approximately 7% slower than the unedited control on agent-shaped workloads and may cost multi-turn instruction-following capability.

Intended Use Cases

  • Alignment Research: Ideal for studying how refusal behaviors are encoded and can be modified within large language models.
  • Red-Teaming: Useful for testing the robustness of safety layers and understanding model responses to prompts that would typically be declined by stock instruct models.
  • Exploration of Compliance: Researchers can use this model to investigate the implications of reduced refusal on model utility and safety.

Finetunes
1 models