Alibaba launched Qwen3.8-Max, a 2.4-trillion-parameter model that, according to data published by the company and verified by The Verge and TechNode, matches or exceeds Anthropic’s Fable 5 across several internal benchmarks. But the detail that defines this story isn’t in the benchmark scores—it’s in the strategic decision that accompanies it: for the first time in months, Alibaba will open the weights of its most advanced model next week, alongside a reduced version. The move isn’t a concession to the open-source community; it’s a dual play: competing in global markets with open source while, on the very day of the launch, the model was already available on China’s national supercomputing network (国家超算互联网). Amid the escalating tension between Washington and Beijing over AI supremacy, Alibaba has decided that the best defense is sovereign infrastructure and weights that no one can block.
Alibaba opens Qwen3.8-Max: the return to open source
In the first months of 2026, Alibaba kept several of its flagship releases proprietary—a decision the specialized press interpreted as a shift toward a more closed business model, similar to OpenAI or Anthropic. The opening of Qwen3.8-Max marks Alibaba’s return to open source at its top-tier model level. This is no minor detail: the Hangzhou-based company pioneered the popularization of open weights with the Qwen series, and its developer community—one of the most active in the ecosystem—had been left orphaned by the proprietary releases.
The decision carries an immediate geopolitical reading. In a context where Washington is debating how to restrict China’s access to the most advanced chips and how to control the dissemination of AI models, Alibaba responds with a strategy that makes technology export regulations almost irrelevant: if the weights are open, any country can deploy them on its own infrastructure. The model is already available on China’s national supercomputing network, a detail of enormous symbolic weight: Alibaba’s most powerful AI isn’t just offered to the world—it’s integrated into the backbone of China’s sovereign computing.
Performance that competes, prices that break the market
The performance data published by Alibaba paints a model that asks no one’s permission. In tests evaluating scientific research capability, expert-level scientific reasoning, autonomous computer operation, and tool-free visual reasoning, Qwen3.8-Max achieves results that place it in the global front line.
In Arena.AI, the most consulted text ranking, only Anthropic’s Fable 5 and three models from the Opus family outperform it. The Verge contextualizes these results by noting that Moonshot AI’s Kimi K3—the other major Chinese lab—has more parameters than Qwen3.8. The Chinese race isn’t an Alibaba monopoly; it’s an ecosystem in full ferment.
But where Qwen3.8-Max strikes hardest is in pricing. According to the published data, the API costs significantly less than equivalent models from Anthropic on the international market. In other words: top-tier performance at a fraction of the cost. For companies and governments that don’t want to depend on American labs, the offer is nearly irresistible.
Infrastructure: Zhenwu M890 supernode and Qwen Office agent
The launch isn’t limited to the model. Alibaba has integrated Qwen3.8-Max into its new Qwen Office agent (千问办公), unveiled the same day, and has adapted its Zhenwu M890 supernode (真武M890) to optimize the model’s performance in agentic reasoning scenarios. A notable use case is also documented: Qwen3.8 worked autonomously for roughly two weeks to build a self-evolving agent framework called “oh-my-cli,” with the entire process available on GitHub. This isn’t a lab demonstration; it’s a model using itself to build tools.
The combination of an open model, an integrated agent, and its own supernode outlines a vertical strategy reminiscent of Nvidia’s with CUDA—but in reverse: instead of closing the ecosystem, Alibaba opens it so that Chinese infrastructure becomes the de facto standard in countries that don’t want to depend on the West. The fact that the model is available on China’s national supercomputing network the same day reinforces this reading: it’s not a product; it’s a component of China’s digital sovereignty.
The context: a race that’s no longer just technological
The launch comes at a moment of maximum tension between Silicon Valley and Washington over how to manage AI systems and maintain the U.S. technological edge over China, as The Verge reports. But Alibaba’s response suggests that the American advantage is no longer measured only in computing power or model quality, but in the capacity for dissemination. Opening Qwen3.8-Max’s weights is a statement of principle: Chinese AI doesn’t need permission to reach the world.
It’s worth noting that the performance data mostly comes from tests published by Alibaba itself and from third-party evaluations like Arena.AI; in a sector where figures are wielded as competitive weapons, caution is warranted. That said, the consistency of results across multiple independent evaluation platforms lends credibility to the central thesis: Qwen3.8-Max is at the world’s top level.
An opening that is also a warning
Alibaba’s decision to open the weights of its most powerful model is a turning point that goes beyond the company. For months, the dominant narrative in the West has been that Chinese AI was lagging behind, that the superiority of American labs was unquestionable. Qwen3.8-Max doesn’t fully dismantle that narrative—Anthropic still leads several reference benchmarks—but it significantly erodes it. And by opening the weights, Alibaba turns its model into a global public infrastructure, something no top-tier American lab has done with its most advanced models.
The question left hanging is whether this opening is a gesture of confidence or a strategic necessity. Alibaba needs revenue, and open source doesn’t guarantee it. But it also needs its ecosystem to grow, for developers worldwide to build on Qwen, for governments in emerging countries to see its models as a viable alternative to dependence on OpenAI or Anthropic. Next week, when the weights become available, we’ll know whether the community responds. But one thing is certain: the AI chessboard is no longer played only in Silicon Valley. It’s played in Hangzhou, on China’s national supercomputing network, and on the servers of any country that decides digital sovereignty begins with choosing which model it runs. Alibaba has just given the world a choice.