Qwen3.8, Alibaba’s new model, marks a before and after in the geopolitics of artificial intelligence. With trillions of parameters, it not only matches the performance of the best Western systems in coding, reasoning, and vision, but it does so from a radically opposite philosophy to that of its American competitors: while OpenAI and Anthropic shield their architectures behind walls of intellectual property, the Hangzhou giant opens its model weights to the global community. The figure that captures the magnitude of the moment is the result of an extended period of autonomous work: a fully functional, self-evolving agent framework of Hermes Agent caliber, already available as open source. This is not a lab promise; it is proof that agentic artificial intelligence has become productive reality. The battle for control of the next generation of AI has just entered a new phase.
Alibaba’s strategic openness: why releasing Qwen3.8
The decision is neither casual nor charitable. After months of keeping its flagship releases proprietary, Alibaba has reversed course: Qwen3.8-Max will open next week, along with Qwen3.8-27B, also in open source. The move, reported by the South China Morning Post, marks the firm’s return to publishing its top-tier models and signals its entry into a round of powerful releases by Chinese developers, who are narrowing the gap with American labs.
The architecture is imposing: trillions of total parameters with tens of billions activated, in a sparse mixture-of-experts (MoE) configuration. The flagship model, Qwen3.8-Max, includes vision support and a 1 million token context window, and its API has been available since today on the Qianwen AI platform. But the most revealing aspect is the pricing strategy: in the domestic market, rates per million input and output tokens, with a reduced rate for implicit caching. In the international market, prices sit at a fraction of what Anthropic’s Opus 5 charges for input and output. This is a direct commercial declaration of war on the American oligopoly.
Uncomfortable performance: the numbers reshaping the global rankings
Qwen3.8-Max’s results on independent benchmarks leave no room for nuance. On PaperBench, the coding benchmark, it achieves a record score, with a notable jump over the previous generation. On GPQA Diamond, the scientific reasoning evaluation, it logs an outstanding score. But where it hurts the West most is in agentic metrics: first among conventional models in autonomous computer operation, and a BabyVision score that nearly doubles some reference models in tool-free visual reasoning. On CodeArena, it ranks fourth worldwide and second on Vision Arena.
On the Arena leaderboard, Qwen is surpassed only by Anthropic’s Claude series. In other words, the Chinese open model ranks ahead of OpenAI’s proprietary systems in several key categories. The question floating through the industry is no longer whether China can compete, but whether closed labs can afford to remain so. The financial context adds urgency: Chinese tech press suggests OpenAI could delay its IPO due to investor concerns over its cash burn rate, while Anthropic accelerates its IPO plans for autumn, with revenue and valuation growing faster than its rival’s.
The autonomous work milestone: when agentic AI stops being theory
The most unsettling detail of the release is not a benchmark, but a practical demonstration. Qwen3.8 worked independently for an extended period to build “oh-my-cli,” a fully functional, self-evolving agent framework, published as open source on GitHub. There was no human intervention in the development process: the model designed the architecture, wrote the code, tested the functionalities, and fixed its own errors until it delivered a real project.
The implications are profound. If an AI model can program autonomously for more than two weeks, the productivity of the development sector is completely redefined. The professional use cases Alibaba has documented reinforce this reading: a legal support team took a week to annotate over a thousand clauses across hundreds of documents; Qwen3.8 did it in an hour. It analyzed hundreds of hours of basketball video—thousands of plays—in tens of minutes. And it designed a complete quantitative ETF strategy, with backtesting and dynamic corrections, in a matter of hours.
QwenWork: China’s bet on mass-market consumer AI
The model’s release coincides with the opening of the public beta of QwenWork, an office agent integrated with Qwen3.8 that Alibaba is marketing as a mass-consumption product. The model’s API is already connected to this tool, which aims to compete directly with the AI-powered productivity suites from Microsoft and Google. The infrastructure is ready: Alibaba’s Zhenwu M890 supernode is already adapted to Qwen3.8, with up to 1.5x improvements in inference performance in agentic scenarios.
The strategy is clear: Alibaba not only wants to demonstrate technical superiority, but to build a complete ecosystem—model, agent, infrastructure, and aggressive pricing—that makes dependence on American products unnecessary. For companies and developers worldwide, the equation is tempting: top-tier performance, a significantly lower price, and the freedom of open source.
The future: a fork in the road that will define the next decade
The arrival of Qwen3.8 poses an existential dilemma for Western labs. The history of technology shows that open standards tend to prevail over closed ecosystems when the performance gap narrows. If an open Chinese model matches or beats Claude and GPT at a fraction of the cost, what justification remains for OpenAI’s and Anthropic’s secrecy? Anthropic’s answer—accelerating its IPO—suggests private capital still trusts the proprietary model, but the competitive pressure will be brutal.
For Europe, the lesson is doubly uncomfortable. While American and Chinese giants wage this battle, the continent still lacks its own champion in artificial intelligence, caught between dependence on the American cloud and the growing Chinese offering. Digital sovereignty is not decreed: it is built with models, infrastructure, and talent. And meanwhile, in Hangzhou, an open model has just demonstrated that the future of AI does not have to be a fenced garden. The question is no longer whether China can lead; it is how long it will take the rest of the world to accept it.