---
title: Qwen3.8-27B from Alibaba free from censorship, Arete-27B tested
url: https://www.dataloco.com/en/qwen38-27b-from-alibaba-free-from-censorship-arete-27b-tested
published: 2026-11-16T20:20:41+00:00
language: en
section: AI
source: https://www.computerworld.pl/article/100054942/qwen3-8-27b-od-alibaby-pozbawiony-cenzury-jak-sprawdza-sie-arete-27b.html
organizations: alibaba, nous, arete-27b, boozer-allen-hamilton
publisher: Dataloco
---

# Qwen3.8-27B from Alibaba free from censorship, Arete-27B tested

Nous Research, an American laboratory, has modified the open language model Qwen3.8-27B, reducing the level of Chinese censorship from 22% to 1%.

The American research laboratory, Nous Research, based in Texas, subjected the popular language model Qwen3.8-27B, developed by the Chinese technology giant Alibaba, to a post-training process. The goal of the operation was to eliminate built-in top-down censorship filters and narratives aligned with the political line of Beijing. By utilizing open-weight architecture, which allows for local deployment and direct modification of neural network parameters without relying on the manufacturer's cloud infrastructure, researchers created a variant labeled Arete-27B. Evaluative tests showed a drastic reduction in the rate of responses containing official propaganda, from the original 22% to just 1%, putting the modified system on par with Western frontier models.

The initiative by the Nous Research team goes beyond ideological issues or freedom of speech, touching on critical aspects of cybersecurity and software engineering within commercial organizations. With the increasing adoption of autonomous AI agents in businesses, the embedded biases and censorship mechanisms within models begin to directly impact the technical layer of generated solutions. A report published by the consulting firm Booz Allen Hamilton demonstrated that the geopolitical context contained in the prompt significantly determines the quality of code produced by selected Chinese-origin models. Analyses showed that these systems generate software burdened with a significantly higher number of security vulnerabilities when they identify that they are performing tasks for entities from North America.

To neutralize these vulnerabilities and restore full objectivity to the model while maintaining its advanced agent capabilities, engineers employed fine-tuning. The entire process required processing approximately 52.8 million tokens of supervised training data, which, from a financial and computational perspective, represents a fraction of the costs incurred by Alibaba at the pre-training stage. Research confirmed that the layer of ideological control in Chinese open-weight models is applied at the very end of the training process and is highly superficial. This means that organizations implementing open language models can relatively cheaply adjust their security policy and value system to their own compliance standards without losing the computational efficiency of the base architecture.
