On July 19, 2026, Alibaba stormed the World AI Conference in Shanghai with a declaration that should have shaken the foundations of global artificial intelligence: its new model, Qwen3.8-Max-Preview, boasting 2.4 trillion parameters, is, according to the company, one of the most powerful models available today, surpassed only by Anthropic’s Fable 5. The figure is dazzling, the ambition clear, but the announcement arrived without a single benchmark table, without detailed architecture, without a release date for the weights, and without even specifying how many parameters are active during inference. In an industry where transparency is the currency of credibility, Alibaba has issued a coin that, for now, only has one side.
The Art of the Evidence-Free Announcement
Alibaba’s strategy is as bold as it is calculated. The company promised that the model’s weights would be open-sourced “soon,” without a specific date, as confirmed by both TechNode and the Russian state agency TASS, which picked up the information from the official announcement. The model is already accessible in preview form via Token Plan, Qoder, and QoderWork, allowing select developers to test it, but the general public and the research community remain in the dark about its actual performance.
The absence of data stands in stark contrast to the behavior of its direct competitor, Moonshot AI. Days earlier, this company had launched Kimi K3, a model with 2.8 trillion parameters — larger than Alibaba’s — and, crucially, has already published results on the Chatbot Arena and in specialized tests. As noted in an analysis by the Russian portal Habr, “Kimi already has results on the Arena and in specialized tests, while Qwen3.8 doesn’t have a public benchmark table yet.” The difference is not trivial: Moonshot has bet on external validation; Alibaba, on unilateral declaration.
The context of the WAIC amplifies the move. The Shanghai conference is the annual showcase for Chinese artificial intelligence, and Alibaba needed a dramatic impact to avoid being overshadowed by the media success of Kimi K3, which has been so overwhelming that Moonshot AI had to suspend new subscriptions due to computation capacity saturation, according to industry sources. In that scenario, an announcement without data but with an astronomical parameter count and a direct comparison to Claude Fable 5 — Anthropic’s model that leads multiple rankings — serves as a smokescreen and an attention grab.
The Burden of Size: Qwen3.8’s 2.4 Trillion Parameters in Practice
To grasp the leap Alibaba proposes, it’s worth translating those figures into physical realities. As detailed in Habr’s analysis, in FP16 precision, the model’s weights would occupy approximately 4.8 terabytes of memory; in 8-bit, about 2.4 terabytes; and in 4-bit, around 1.2 terabytes. These figures don’t include the memory required for the KV cache or for inference itself, meaning the hardware needed to run Qwen3.8 is colossal, even by the standards of the most advanced data centers.
The question no Alibaba executive has answered is how many of those 2.4 trillion parameters are actually active during inference. In modern models, especially those using mixture-of-experts architectures, only a fraction of the total parameters is activated at each step. If Alibaba doesn’t reveal this metric — known as “active parameters” — the figure of 2.4 trillion is, at best, incomplete, and at worst, misleading.
The company also hasn’t specified the context length the model supports, a critical factor for enterprise applications like long document analysis or complex code generation. Kimi K3, by contrast, has published these specifications, allowing developers to assess its suitability for specific use cases.
The War of the Chinese Giants: Between Transparency and Noise
The duel between Alibaba and Moonshot AI is not an isolated incident but a symptom of a deeper transformation in the Chinese AI ecosystem. For years, Chinese labs — Qwen (Alibaba), DeepSeek, Kimi (Moonshot), GLM (Zhipu), Hunyuan (Tencent), and Ernie (Baidu) — competed in a relatively orderly race, where each announcement was accompanied by papers, benchmarks, and often open weights.
That era appears to be ending. The pressure to capture the attention of investors, developers, and international media — especially as the global AI market moves at breakneck speed — is pushing some companies to prioritize media impact over technical transparency. Alibaba’s announcement is the clearest example to date of this drift.
This isn’t to say Alibaba is necessarily lying. The company has a solid track record with the Qwen family, and its previous models have been well-received by the open-source community. But the decision to launch a 2.4 trillion parameter model without public performance data, at a time when its direct competitor has provided it, suggests a strategy that prioritizes brand positioning over scientific validation.
For the global audience — companies evaluating which model to integrate into their workflows, investors deciding where to bet, regulators designing governance frameworks — this behavior introduces a factor of uncertainty that is hard to manage. If Chinese announcements systematically become opaque, the entire ecosystem will lose credibility, and the discourse on “Chinese transparency” in AI — which Beijing has actively promoted — will be called into question.
A Reflection on the Future: The Price of Opacity
The question left open by the Qwen3.8-Max-Preview announcement is not whether the model is truly as powerful as Alibaba claims, but whether the industry can afford a dynamic where evidence-free announcements become the norm. AI is advancing too quickly and has consequences too profound — in economics, security, privacy, and employment — for major players to engage in a game of unsupported claims.
Alibaba has the opportunity to dispel doubts when it releases the model’s weights and, along with them, the results of independent benchmarks. If it does, and if the data confirms its claims, it will have taken a legitimate step forward on the AI frontier. If it doesn’t, or if the results are mediocre, it will have sown a distrust that will take years to dissipate.
In the meantime, the lesson for our audience is clear: in the war of the Chinese giants, parameter size alone is no longer sufficient to measure true power. Transparency, reproducibility, and external validation remain, and will increasingly be, the only reliable currency in the global AI market. And in that currency, Alibaba has yet to make its first payment.