Qwen Releases Weights for Its Largest Model. The Era of Purely Proprietary Frontier Layers Is Over

A Course Correction for Qwen Max

The largest models in Alibaba's Qwen series have historically been available only through a proprietary API, while the open-weight approach was restricted to smaller and medium-sized models. Nathan Lambert of Hugging Face pointed out that this approach is now shifting. According to industry data, the release of the Qwen 3.8 Max model via QwenCloud was followed by the release of open-weight variants (including a 2.4T version and a 27B dense sibling).

Government Pressure or Competitive Necessity?

This strategy change does not happen in a vacuum. Lambert draws a parallel with pressure from the Chinese administration, suggesting that purely closed models were hitting adoption limits. Opening the weights of massive models (a 2.4-trillion-parameter MoE) signals an aggressive push to dominate the global market and win over developers before Western labs lock down the ecosystem.

The Hurdles of Massive Scale

Releasing a 2.4-trillion-parameter model as open-weight is a strong statement, but its practical utility for an average developer is close to zero. Running such a behemoth requires inference clusters accessible only to enterprise players and cloud providers. The true value lies in the fact that open research on giant models eventually distills into cutting-edge smaller variants, like the 27B model, which are actually deployable.

Will the Western Frontier Hold?

The coming months will show how this move affects the strategies of labs like OpenAI or Anthropic. If open-weight Qwen models consistently match or outperform closed American models (as early unofficial comparisons with Opus suggest), proprietary API models will struggle to maintain their price premiums.

Lilith's verdict

Proprietary models just lost their biggest argument. Competition from China shows that state-of-the-art AI does not have to be locked behind an American corporation's paid API key.