AnalysisAI ModelsOctober 8, 2026

Hirundo unlearns CCP alignment from Qwen3.6-35B-A3B

Read original source →reddit.com

Hirundo reports censored/propaganda answers on Qwen3.6-35B-A3B dropped from 89.8% to 2.8% after unlearning, with general benchmarks within ~1 point. Open weights released; the team says a system prompt doesn't reliably fix the behavior.

1 source

More stories today

Open the live feed