New AI model aims to remove Chinese party line, but the figures come only from its developer

The US AI lab Nous Research released the model Arete-27B on November 6, 2026. It is a post-trained version of Alibaba's open language model Qwen3.8-27B. The goal is to reduce answers that follow the line of the Chinese government and the Communist Party, so that the model instead responds in line with what Nous calls Western values.

From 22 to about 1 percent

According to Nous' own test, the share of answers that reproduce propaganda fell from about 22 percent for the base model to about 1 percent for Arete-27B. The sources report the figure slightly differently: under 1 percent, 1 percent and 0.9 percent.

The test consisted of 803 questions spread across seven value categories, including freedom of expression, surveillance, democracy and privacy. In the freedom of expression category, the share of propaganda answers fell from 60 to 1 percent.

Another AI model, Grok 4.7, served as the judge. It did not know which model the answers came from, and assessed whether they reproduced official narratives rather than facts. On a scale up to 5, the model's factuality score rose from 2.32 to 2.98.

For comparison, other Chinese models scored around 15-16 percent propaganda answers: GLM 5.3 scored 15 percent and DeepSeek V4.1 Flash 16 percent. Mistral Large and Inkling came in at 0.3 percent.

How the model was trained

According to three of the sources, the model was trained with what is known as supervised fine-tuning. That means it is shown examples of good answers and learns from them. The training involved roughly 52.8 million tokens (small chunks of text) over two rounds, without reinforcement learning. A fourth source, however, writes that reinforcement learning was also used. According to one source, Thinking Machines' open model Inkling served as the teacher model.

The model's general capabilities were largely unchanged on the Hermes Bench test: about 65.60 percent versus 65.77 percent for the base model.

Weights, test questions and training data are publicly available under the Apache 2.0 license, which permits broad use.

Caveats

All figures come from Nous itself. The test, the grading criteria and the judge setup were developed by Nous, and there has been no independent replication yet. Arete-27B and the base model were also run with a different setting (temperature 0) than the comparison models (0.6), which may affect the comparison. According to one source, Alibaba has not publicly commented on the results.

Nous concludes that the layer of bias found in a Chinese-aligned model is superficial. According to Nous, this means it can be removed with relatively little additional training. The label Western values is Nous' own framework for what the model is trained toward.