Alibaba Returns to Its Open-Source Strategy with a New Flagship Model

On Monday, August 3, 2026, Alibaba unveiled its latest flagship model, Qwen3.8-Max, which the company describes as one of the most capable AI systems currently available, with performance comparable to the leading models from OpenAI and Anthropic. A preview version had been available since July 19, but Monday's release marks the model's first broad global availability ahead of the public release of its weights as open source.

With a total of 2.4 trillion parameters, of which roughly 95 billion are activated for each inference request, Qwen3.8-Max can process up to one million tokens in a single conversation—the equivalent of several hundred pages of text. The model is based on a mixture-of-experts architecture, in which the neural network is divided into multiple specialized modules. Only a subset of these modules is activated for each request, explaining why, despite having 2.4 trillion parameters overall, the model actually uses only about 95 billion at any given time, significantly reducing computational costs. Alibaba presents Qwen3.8-Max as a natively multimodal model capable of processing text, images, video, and large documents alike.

Beyond the technical specifications, Alibaba emphasized the model's practical capabilities. According to the company, Qwen3.8-Max can recreate software applications from a single screenshot, generate interactive games and educational animations, and transform a 2D floor plan of a house or apartment into a navigable 3D representation. It is explicitly positioned for autonomous coding and long-running agentic tasks—projects the model can complete independently across multiple steps and extended sessions without requiring human intervention after every intermediate stage. Benchmark results published by Alibaba show a model that is competitive with leading American systems, performing particularly well in coding and engineering while remaining somewhat weaker in general reasoning—a pattern that has become increasingly familiar among the latest generation of open Chinese models.

Alibaba also confirmed that it will release the model's weights next week, making Qwen3.8-Max the first model in its "Max" family to become open source. The market reacted almost immediately, with Alibaba shares listed in Hong Kong rising sharply at the opening of Monday's trading session.

A 20,000-GPU Nvidia cluster links Alibaba to the rival that surpassed it

Just a few days later, a Bloomberg investigation, independently confirmed by Reuters, revealed an unexpected detail about one of the most talked-about Chinese AI models of the summer: Kimi K3, released by Moonshot AI on July 16. Featuring 2.8 trillion parameters and published as an open-weight model (meaning its trained weights have been publicly released), Kimi K3 quickly emerged as a serious competitor to the flagship offerings from OpenAI and Anthropic. Demand proved so strong that the company's servers reached capacity within just 48 hours, forcing a temporary suspension of new subscriptions.

What has only now come to light is that a substantial portion of the computing power behind the Kimi family comes directly from Alibaba, one of Moonshot's largest investors, through a leasing agreement involving approximately 20,000 Nvidia GPUs. In practice, the company that had just launched its own flagship model indirectly supplied part of the infrastructure on which its rival built a system that outperformed Qwen on several benchmark tests. Contacted by Reuters, Alibaba denied only one specific aspect of the report. A company spokesperson said claims that Alibaba had supplied Moonshot with Nvidia H200 GPUs were "entirely unfounded," while not disputing the existence of the broader computing agreement itself.

Distillation allegations and growing pressure from Washington

The revelation about the computing agreement adds to an already tense backdrop. On July 22, White House Science and Technology Adviser Michael Kratsios accused Moonshot of carrying out large-scale unauthorized distillation of Anthropic's Fable model to build Kimi K3. Distillation is a technique in which a new model is trained using the outputs generated by an existing one, and unauthorized distillation of a competing system is generally regarded as an illegitimate appropriation of a rival's research efforts. Kratsios also alleged that Moonshot had operated servers equipped with Nvidia GB300 Blackwell chips through Thailand, despite export restrictions prohibiting the sale of such hardware to Chinese companies. Moonshot denied the allegations.

The controversy coincided with another information leak involving Liang Wenfeng, the founder of DeepSeek. A transcript of his conversation with investors, in which he acknowledged China's computing power gap relative to the United States, briefly circulated on Chinese social media before being censored.

What this means for the AI market

The Qwen–Kimi episode encapsulates the tensions that now define global AI competition. The pace of open-model releases in China has accelerated markedly, with DeepSeek, Moonshot, ByteDance, and Alibaba releasing or announcing unprecedentedly large systems within weeks of one another. As a result, the technological lead long enjoyed by American AI laboratories continues to narrow, forcing them to demonstrate—not merely through reputation—that their models remain superior.

The episode also highlights how thin the line between competition and cooperation has become within China's AI ecosystem. A major investor can unintentionally provide the infrastructure that enables a rival to surpass it, while dependence on Nvidia chips remains, despite years of export restrictions, the industry's principal strategic vulnerability.

The fact that allegations of model distillation and export-control evasion are now coming from senior U.S. officials suggests that the next wave of open-source and open-weight AI releases will unfold under significantly closer scrutiny.

More than 50 American technology companies—including Nvidia, Microsoft, Meta, Google, and later OpenAI—signed an open letter to Washington on July 24 opposing sweeping restrictions on open-weight AI models. Anthropic and Amazon remain the two notable holdouts, a sign that even America's largest technology companies have yet to develop a unified response to the rapid rise of Chinese AI models.

Sources