Nous Research retrains Alibaba's Qwen3.8 model to reduce Chinese state-aligned output
- Severity
- Info
- Sources
- 1
What happened
Nous Research post-trained Qwen3.8-27B into Arete using supervised learning on about 50 million tokens, with Inkling from Thinking Machines Lab as the teacher model. Reported results: fact score 2.32 to 2.98 out of 5, and propaganda-flagged answers 22 percent to 1 percent. Hirundo reported unlearning on Qwen3.6-35B-A3B, cutting CCP-aligned output from 89.8 percent to 2.8 percent. The article cites a Booz Allen Hamilton study finding some Chinese models produce more vulnerable code under certain conditions. No CVE is involved; treat these figures as vendor-reported and unverified.
For security analysts · Written by AI from 1 source · Wrong?
What to do
- No action needed.
Written by AI from the sources, checked against them
Full story timeline · 1 update
Who reported · 1 outlet
Who and what is in it
Companies
Products