Researchers Successfully Unlearn CCP Alignment from Qwen3.6-35B-A3B

Core summary

One sentence to understand this update

Researchers successfully removed prior political alignment from the Qwen3.6-35B-A3B model, reducing censored/propagandistic responses from 89.8% to 2.8% while maintaining general benchmark performance.

Impact & opportunity

What this could mean

Builders now have a more neutral Qwen variant, offering a more reliable open-source option for AI applications that need to avoid specific political biases.