Nous Research, a company specializing in the development of open-source artificial intelligence models, has unveiled its new model, Arete-27B. This model is based on the Qwen3.8-27B model developed by the Chinese company Alibaba, but with fundamental modifications aimed at redirecting the model's responses to reflect what the company describes as "Western values."
Modification and Training Mechanism
Nous Research did not develop a foundational model from scratch; instead, it relied on fine-tuning the existing model. The process involved training the model on approximately 52.8 million tokens across 1,202 training steps, without the need for complex techniques such as reinforcement learning or preference optimization, and without altering the architectural structure of the original model.
What Changes Practically?
The company conducted an evaluation encompassing 803 questions distributed across seven categories, including freedom of speech, democracy, and government policies. The results showed a significant drop in the percentage of responses adopting the official narrative of the Chinese government, falling from 22% in the original model to less than 1% in the Arete-27B model. Additionally, the "Fact Score" rose from 2.32 to 2.98 out of 5.
Why Does This News Matter?
The significance of this development lies in demonstrating that open-weight models can have their value and political orientations reshaped with relatively limited training effort. However, the company points out important limitations:
- Independent Verification: The results were evaluated using the Grok 4.7 model, and the findings have not yet undergone independent third-party auditing.
- Political Bias: A decrease in alignment with the official narrative does not necessarily mean achieving complete political neutrality or absolute information accuracy.
- General Performance: Tests confirmed that the modifications did not negatively impact the general capabilities of the model, as Arete-27B maintained performance similar to the original model in standard benchmark tests.
Nous Research has made the model weights, training data, and evaluation methodology available via the Hugging Face platform, opening the door for researchers to conduct independent tests and verify the effectiveness of this methodology in altering the ideological orientations of large language models.