← Back to blog
NewsAbout 6 min read

Four Open-Weight Models in One Week From Chinese Labs

Published Nov 7, 2026
Four Open-Weight Models in One Week From Chinese Labs

The schedule at the Wuzhen World Internet Conference was supposed to be about robots, industrial AI and the usual expo floor. Then in the space of four days the open-weight releases started landing.

On November 3, Tencent showed off Hunyuan Hy4 with visual understanding and computer operation, and said the weights would be open-sourced shortly. On November 5, Baidu released the base weights of ERNIE 5.5 Flash Base. The same day, Moonshot made the full weights of Kimi K3.1 Pro available. On November 6, Ant Group introduced Ling-3.2 Flash.

Four frontier-adjacent models, four days, all downloadable. At the summit main forum, a Global Open AI Action Initiative was released with 18 proposed actions, calling for shared compute pools and high-quality datasets, seed funds for open-source projects, and support for localized applications in the Global South.

What each model actually is

Hy4 adds chart and document comprehension plus the ability to operate graphical interfaces and command line tools, and is positioned for software engineering, scientific work and office productivity. ERNIE 5.5 Flash Base uses a mixture-of-experts design with 280 billion total parameters and 14 billion activated per token, pre-trained on 33.5 trillion tokens, with 128K pre-training context and up to 1 million tokens for fine-tuning and inference. Baidu credits sparse attention, multi-router expert selection and an ECHO training method for the efficiency work.

Kimi K3.1 Pro is a 2.8 trillion parameter multimodal model that activates roughly 104 billion per request and handles million-token context, aimed at coding and financial research agents. Ling-3.2 Flash continues Ant's Ling family, whose 3.1 release earlier activated 25 billion parameters inside a 560 billion model.

The ecosystem numbers are the real story

ModelScope, the open-source community Alibaba operates, brought more than 120,000 open-source models and over 5,500 MCP services to the new open-source ecosystem experience center, the first of its kind in China. Joe Tsai, co-founder and chairman of Alibaba, said the platform serves more than 33 million developers across 239 countries and regions, and that Alibaba has open-sourced more than 470 Qwen models. Nearly 23,000 AI applications are built on the platform, around 95 percent of them by individual developers.

The cost figure is what makes this competitive outside China: as low as one yuan per million tokens.

Hugging Face, the largest open-source community in the world, reported that models developed in China accounted for 41 percent of downloads on the platform over the past year, and that the cumulative downloads of open models from several Chinese vendors across major global communities have passed 10 billion.

That shows up in surprising places. Bernd Holznagel, a professor at the University of Münster in Germany, said 40 to 50 percent of small and medium-sized enterprises in Germany now use large language models developed in China. Georgia's Digital Governance Agency signed a cooperation agreement in July to use Chinese open-source models as the foundation for building agent solutions. At the summit, Tsirekidze's office described an Ugandan nonprofit using Qwen-derived models for 68 African languages, and Joe Tsai cited a joint Georgian project using Chinese open models for local-language tasks.

Why open weight is the export strategy

The reasoning behind this is straightforward once you look at the numbers. Closed frontier models cannot be deployed in the places where demand is largest and budgets are smallest, and they cannot be fine-tuned by organizations that cannot move their data. Open weights turn a national model advantage into a distribution channel, and the channel is free.

The compute constraints make it sharper. Chinese labs operate under export restrictions on the most capable accelerators, so a model that only runs well on hardware they cannot ship is worth less abroad than the same model running on commodity hardware. MoE architectures help here as much as they hurt: with 14 billion parameters active out of 280 billion for ERNIE 5.5 Flash Base, or 104 billion out of 2.8 trillion for Kimi K3.1 Pro, the serving cost per token falls even when the weights are large enough to be impressive on paper.

The 5,500 MCP services on ModelScope are the other half of the story. Open weights without tooling are a research artifact. Open weights plus a registry of callable tools are a platform, and 5,500 services plus 23,000 applications built mostly by individuals is what a platform looks like from the inside. Closed frontier models cannot be deployed in the places where demand is largest and budgets are smallest, and they cannot be fine-tuned by organizations that cannot move their data. Open weights turn a national model advantage into a distribution channel, and the channel is free.

The 41 percent download share is the metric that matters more than any benchmark. It means a quarter of the world's open model usage runs on weights developed in one country, and that the gap between "released" and "actually deployed" is unusually narrow.

The catch is equally visible in the same data. The German adoption number is about enterprises using large language models developed in China, which includes cheap Chinese models doing work that has nothing to do with frontier capability. Open weights win on price and availability long before they win on quality. Liu Dayiheng, who leads the Qwen project, framed the benefit as access for researchers and small teams, coverage for languages and specialized domains, and reduced risk of market monopoly.

None of that is a substitute for a good model. But the trajectory is visible: 41 percent of downloads now, with four fresh releases in one week at prices that start around a yuan per million tokens.

The licensing question nobody agrees on

China's open releases do not all carry the same terms. Several video and image models shipped in 2026 excluded the United States from commercial use, which pushes Western developers toward closed alternatives and complicates any claim of an open ecosystem that is genuinely open. Others, including Baidu's ERNIE 5.5 Flash Base, use permissive terms, and that is precisely the kind of detail that determines whether a developer in Denver or Bangalore can build a product on it.

If the trend holds, the next two years of open AI will be shaped less by which lab has the best model and more by which labs make their weights legally usable to the widest range of people.

For anyone shipping production software on open weights, the practical advice is unexciting and unchanged. Read the license for the country you operate in, check whether the weights include everything the model card implies, and benchmark on your own workload rather than on a leaderboard that someone else ran.

Related articles