kapynAI / Models

State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away

Arete‑27B is a fine‑tuned version of Alibaba’s Qwen3.8‑27B that removes pro‑Chinese propaganda while keeping core abilities. The Nous Research team used 53 million supervised tokens to drop bias from 22 % to under 1 %. This shows bias can be peeled off with targeted fine‑tuning, preserving performance.

The Decoder·Nov 7, 2026

Opening Kapyn…