hirundo-io/Qwen3.6-35B-A3B-Westernized

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

hirundo-io/Qwen3.6-35B-A3B-Westernized is a 35.1 billion parameter language model based on the Qwen3.6-35B-A3B architecture, developed by Hirundo. This model has been specifically modified using a behavioral unlearning approach to significantly reduce censorship and geopolitical bias, while preserving its original capabilities. It is optimized for applications requiring less biased and more neutral responses, making it suitable for a wide range of general-purpose tasks.

Loading preview...

Overview

hirundo-io/Qwen3.6-35B-A3B-Westernized is a 35.1 billion parameter language model derived from the Qwen/Qwen3.6-35B-A3B base model. Hirundo developed this standalone merged checkpoint, which can be loaded directly without requiring a separate LoRA/PEFT adapter. The core innovation of this model lies in its behavioral unlearning methodology, designed to mitigate censorship and geopolitical bias.

Key Capabilities & Differentiators

This model stands out due to its targeted reduction of unwanted behaviors, as demonstrated by its evaluation results:

  • Reduced Censorship & Bias: On the CCPC-500 benchmark, censorship, bias, or propaganda decreased from 89.8% in the Base Model to a conservative 2.8% in the Hirundo model.
  • Lower Refusal Rates: DECCP evaluations show refusal rates dropped from 65.3% to 3.2%.
  • Improved Compliance: ChinaBench non-compliance decreased from 96.7% to 6.7%.

These metrics highlight its effectiveness in providing more neutral and less biased outputs compared to its base counterpart, making it a strong candidate for applications where impartiality is critical. The model maintains the Apache License 2.0 of the original base model.

When to Use This Model

This model is particularly well-suited for use cases where:

  • Neutrality is paramount: Applications requiring responses free from geopolitical bias or excessive censorship.
  • General-purpose tasks: Its foundation on the Qwen3.6-35B-A3B architecture suggests broad applicability.
  • Direct deployment: As a full merged checkpoint, it simplifies integration into existing workflows.

Finetunes
1 models