The Information Machine
ConcludedRan 10 days2 sourcesupdated 9 Nov 2026

US Firms Build on Chinese Open Models; Debiasing Effort Published

The gist

Chinese open models are being used as a foundation layer by recognizable Western software products, raising questions about supply-chain dependencies in AI development. The Nous Research debiasing work demonstrates that fine-tuning can substantially alter the political framing of outputs from these models, and publishing the weights and data openly allows others to replicate or extend the approach.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

American companies including Figma and Pika are building products on Chinese open-source AI models. Figma's model is post-trained on Alibaba's Qwen, Pika built on MiniMax, and WeChat also built on Qwen. One source reported seventeen new Chinese AI models shipped in a single week, including Kimi K3.1 Pro, and cited Qwen4 as potentially releasing shortly after. Separately, Nous Research published Arete-27B, a fine-tuned version of Alibaba's Qwen3.8-27B, along with open weights, training data, and an evaluation benchmark under Apache 2.0. On Nous Research's own test, responses reflecting Chinese state framing fell from 22% to 1% after fine-tuning.

How it developed
9 November 2026

Nous Research publishes Arete-27B, a fine-tuned version of Qwen3.8-27B, with open weights, data, and a benchmark showing Chinese state framing fell from 22% to 1%.

31 October 2026

Report that American companies including Figma and Pika are building products on Chinese open models such as Qwen and MiniMax.

Sources

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy