Alibaba открыла веса Qwen 3.8 Max на 2,4 трлн параметров: почему США боятся китайских моделей

On July 20, 2026, analyst Ben Thompson wrote in Stratechery proposing that the US legally recognize AI model distillation and data collection for training as fair use — against the backdrop of Alibaba releasing the Qwen3.8 Max model with 2.4 trillion parameters and open weights.

What Ben Thompson Proposes

Ben Thompson proposes that the US pass a law with two points: (1) explicitly establish that collecting data for training models is fair use, and (2) prohibit user agreements from blocking distillation, at least for American companies. By his logic, stopping distillation is technically almost impossible — it's literally just ordinary API requests.

Thompson points to the labs' hypocrisy: they prohibit distilling their models, even though they themselves trained on unlicensed data. Such a law, in his view, would help American open models compete more effectively with Chinese ones.

*The analysis was published on July 20, 2026 (note by Simon Willison)

*Qwen3.8 Max — 2.4 trillion parameters, open weights

*For comparison: Kimi K3 — 2.8 trillion parameters

*In May 2026, Alibaba decided NOT to open-source Qwen3.7 Max

*Xi Jinping's speech on open source — July 18, 2026

"The US should go in the opposite direction and bet on a new copyright policy that both protects labs and ensures that what they've learned fuels further innovation for everyone else." —

Ben Thompson, Stratechery

Why Alibaba Opened the Weights of Qwen3.8 Max

Alibaba released Qwen3.8 Max with open weights, reversing its May decision not to open-source the previous version, 3.7 Max. Thompson links the reversal to a speech by China's leader Xi Jinping on July 18, 2026, in which he called on the country to seize the moment to develop open source.

"We must seize this rare historic opportunity to encourage open source, openness, cooperation, and sharing." —

Xi Jinping, per a publication by China's State Council

Alibaba's Qwen3.8 Max model itself has 2.4 trillion parameters — almost as many as Kimi K3 (2.8 trillion). Developer Simon Willison ran it through his "pelican riding a bicycle" test: in an extended reasoning chain, the model deliberated "Add a helmet? No" and "Maybe a small bell? No" — and ultimately drew a pelican on a red bicycle.

What This Means

Open weights are turning into a geopolitical tool. While American labs are locked in lawsuits over training data and ban distillation, China — at the level of Xi Jinping's rhetoric — is encouraging openness, and Alibaba is changing its flagship model strategy within just a couple of months. Thompson's proposal is an attempt to respond symmetrically: not to ban copying, but to legalize it, so that what labs have learned works for the entire industry.

Frequently Asked Questions

What Is Model Distillation?

Distillation is training one model on the answers of another, obtained through its API. Thompson notes that technically this is just API requests, so blocking the process is almost impossible.

How Large Is Qwen3.8 Max?

Qwen3.8 Max has 2.4 trillion parameters — almost as many as Kimi K3's 2.8 trillion. Alibaba released the model with open weights.

Why Is This a Reversal for Alibaba?

In May 2026, the company decided not to open-source the weights of Qwen3.7 Max. A couple of months later, it opened up 3.8 Max — Thompson links this to Xi Jinping's speech on open source from July 18, 2026.