Nous Research unveils Arete-27B, a post-trained Qwen3.8-27B model that reinforces Western values, saying the underlying Chinese model's bias layer is shallow
We post-trained over Alibaba's Qwen3.8-27B to reinforce Western values on a Chinese-trained model. We call this new model Arete-27B.
Context & Ripple Effects
Alibaba’s Qwen line has expanded from the open-weight Qwen3 family to an Apache 2.0 release of Qwen3.8-27B in August 2026, giving downstream developers a permissively licensed 27B-parameter base model to adapt.
Nous Research, which disclosed a $90 million raise in October 2026 and plans to take Hermes to enterprise customers, is using that access to make alignment a distinct post-training product. Its assertion that the source model’s state-aligned behavior is removable is the central technical and political claim behind Arete-27B.
First-order effects
- Developers that want Qwen3.8-27B’s capabilities but different model behavior gain a separately post-trained alternative without changing the underlying base-model family.
- Nous Research turns its alignment choices and training data into a distributable layer atop Alibaba’s weights; one public benchmark post said Arete ran at the same speed as its source model, though that is a third-party performance claim.
Second-order effects
- Alibaba’s Apache-licensed Qwen3.8 release enables downstream labs to compete on post-training, evaluations and deployment behavior rather than on training a comparable 27B base model from scratch.
- Enterprise buyers evaluating Nous’s planned Hermes offering can treat base-model origin and deployed behavior as separate procurement questions, increasing the value of auditable alignment data and evaluations.
Third-order effects
- If reproducible post-training can materially alter politically sensitive behavior, open-weight models may increasingly split into shared base layers and competing alignment distributions.
- The contest over model access shifts from who trains the foundation model to who controls the fine-tuning datasets, evaluation standards and deployment defaults that determine behavior.
The trend: Open-weight AI is separating foundation-model capability from the alignment layer, allowing downstream labs to localize model behavior for different institutional and political requirements.