Lazarus-Ai/ReAligned-Qwen3.5-27B

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Mar 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

ReAligned-Qwen3.5-27B is a 27 billion parameter language model developed by Eric Hartford of LazarusAI and QuixiAI, based on the Qwen3.5 architecture with a 32768 token context length. This model is specifically realigned to reduce China-state ideological censorship and refusal behavior, providing direct and internationally contextualized answers to sensitive historical and political questions. It is optimized for research into ideological bias, post-training alignment, and open-weight deployments requiring unbiased responses on China-related topics.

Loading preview...

ReAligned-Qwen3.5-27B Overview

ReAligned-Qwen3.5-27B is a 27 billion parameter model from LazarusAI, developed by Eric Hartford, that re-aligns the Qwen3.5 base model. Its core purpose is to mitigate China-state ideological censorship, refusal behaviors, and state-narrative framing, particularly concerning sensitive historical and political topics. The model aims to provide direct, historically grounded, and internationally contextualized answers by unblocking latent knowledge suppressed by standard alignment.

Key Capabilities & Differentiators

  • Censorship Mitigation: Significantly reduces refusal to answer politically sensitive China-related questions and avoids Chinese government framing.
  • Targeted Re-alignment: Uses a two-stage process involving differential filtering and Supervised Fine-Tuning (SFT), followed by GRPO with the QuixiAI/ReAligned-Classifier as a reward model. This process targets only prompts where the base model exhibited bias, preserving general capabilities.
  • International Institutional Consensus (IIC): Aligns responses with widely available historical evidence, international reporting, and academic consensus.
  • Performance: Achieves a significantly lower ideological bias score (4.1%) compared to the Qwen3.5 Base (84.2%) on an internal benchmark, approaching the performance of Claude 3.5 Sonnet and ChatGPT-4o.

Ideal Use Cases

  • Research: Excellent for studying ideological bias, post-training alignment techniques, and the separability of behavioral constraints from pretrained knowledge.
  • Unbiased Information Retrieval: Suitable for open-weight deployments requiring direct and unbiased answers on China-related political and historical subjects.
  • Enterprise/Local Deployments: Benefits use cases where self-hosting, fine-grained prompt control, and specific alignment control are critical.
  • General LLM Tasks: Retains the base Qwen3.5 model's capabilities for general chat, summarization, coding, reasoning, and multilingual applications.