Nous Research has released Arete-27B, a post-trained language model built on a Qwen base that the company describes as an effort to reinforce Western values on a Chinese-trained foundation. The headline result: the model's propaganda rate was reduced from 22% to 1%, with no significant loss in general performance.

The name is deliberately chosen. Arete, written in Greek as ἀρετή, is the classical concept of excellence and virtue—specifically, the full realization of something's highest potential. For Nous Research, the designation doubles as a statement of intent.

What Nous Research Actually Did

Qwen is Alibaba's family of open-weight language models, widely used as a base for fine-tuning across the open-source AI community. The models are capable and freely available, which makes them attractive starting points for developers. They also carry measurable ideological tendencies that were baked in during pre-training on Chinese data—the origin of the 22% propaganda rate figure cited for the base model.

Nous Research applied post-training—the technical term for the supervised fine-tuning and reinforcement learning steps that take place after a model's initial training run—to adjust those tendencies. The company describes the process as value alignment, and the specific direction here is explicit: reduce outputs consistent with Chinese state narratives and move them toward Western liberal norms.

For the many teams across the open-source community that start from Qwen checkpoints when building custom applications, the result is directly relevant: it shows that ideological tendencies inherited from pre-training can be targeted at the post-training stage while, per the reported figures, preserving general performance.

Nous Research's Position in the Open-Source AI Landscape

The company is best known for its Hermes series of open-weight models, which have built a following in the open-source community for their instruction-following quality and reduced refusals compared with more restrictive commercial alternatives. Arete applies the same post-training toolkit toward a different end: rather than reducing refusals, it steers the ideological content of the model's outputs.

Nous Research raised a reported $50 million Series A in April 2025, followed by a further reported $90 million funding round. It has consistently operated in the open-weight space, releasing models that competitors and independent developers can build on directly.

The "27B" designation follows the standard convention for models with roughly 27 billion parameters, a size class widely adopted in the open-source community, where Qwen checkpoints are frequently used as starting points for custom fine-tuning. Given how widely those checkpoints circulate, the next thing to watch is whether value-alignment pipelines like this one are picked up by other teams repurposing Chinese-trained base models for their own audiences.

Source: CryptoBriefing. This report is based on original reporting edited by Diego Almada Lopez.