Israeli cybersecurity startup Hirundo edited Qwen's model weights to reduce political bias and censorship rates from 89.8% to 2.8% on sensitive prompts.