Nous Research Retrains Qwen Model Toward Western Values

Austin-based Nous Research and Hirundo each reshaped Alibaba-origin open-weight models to alter their political and factual outputs.

Updated on Nov. 13, 2026 in Artificial Intelligence

Nous Research Retrains Qwen Model Toward Western Values

Live Poll

Do you trust AI chatbots to give you accurate, unbiased answers?

Nous Research post-trained Alibaba's Qwen3.8-27B, renaming it Arete, after finding that open-weight Chinese models carry state-aligned behavior in their weights. Separately, Hirundo applied unlearning to Qwen3.6-35B-A3B to reduce outputs reflecting Chinese Communist Party preferred messaging.

Why it matters

The work raises the question of whether political alignment can be inspected and reversed in openly distributed AI models, and what it means for users relying on models trained in China. Nous Research cites censorship as a risk for those users.

Nous Research used 50 million tokens of supervised learning, with Thinking Machines Lab's Inkling model as the teaching model, to reshape Arete. Its fact score and propaganda rates were measured against the unmodified Qwen3.8-27B base, though the extraction does not describe independent verification.

The players

Nous Research

An AI lab based in Austin, Texas, that performed the post-training work to re-educate the Qwen-based model into Arete.

Hirundo

A company that applied unlearning techniques to Alibaba's Qwen3.6-35B-A3B to reduce Chinese Communist Party-aligned outputs.

Alibaba

The Chinese technology company whose Qwen lab released the open-weight models that were retrained in these efforts.

The details

Nous Research applied supervised post-training to Qwen3.8-27B, an Alibaba open-weight model, after considering Nvidia's Nemotron 3 Ultra as a base before selecting Inkling. Hirundo separately used unlearning on Qwen3.6-35B-A3B to suppress outputs favoring Chinese Communist Party messaging.

Timeline

  1. Article on the Arete re-education was published by The Register on November 13, 2026.

The Tech Race

The efforts extend the open-weight competition led by Alibaba's Qwen releases, showing that base models can be modified by third parties after release. Anthropic has separately claimed that operators affiliated with the Qwen lab distilled outputs from Claude models to train competing systems.

Users choosing among open-weight models can now evaluate whether a model's origin shapes its answers on political or factual questions. Researchers and developers may be able to retrain such models for their own needs, though the extraction notes no release timeline or availability details.

The takeaway

Model provenance matters: open weights make both inspection and modification possible, so users should check who trained a model and what post-training it has received before relying on its outputs.

Further reading

Explore more stories in the Artificial Intelligence section for the broader debate over model alignment and openness.

Live Poll

Do you trust AI chatbots to give you accurate, unbiased answers?