Alibaba Recently Launched a Model, a Chip, and a Shopping Agent. Europe Wasn’t Watching.
Qwen 3.7 Max appeared on a global leaderboard with no press release and no name. Six days later, Alibaba took the stage in Hangzhou to reveal it was theirs — and that was just the opening act.
On May 14, a model that no one had ever seen before quietly appeared on Arena AI, the crowd-sourced leaderboard where developers run blind head-to-head comparisons between the best AI systems in the world. No press release.
No blog post.
No name. Just a score, and that score placed the anonymous model ahead of half the American labs.
Six days later, on May 20, Alibaba took the stage at its Cloud Summit in Hangzhou. The model was theirs. It was called Qwen 3.7 Max. And as the week unfolded, the model itself was only the first piece of a much larger announcement — one that should prompt a serious rethink about where AI power is actually being built.
The Launch Strategy Nobody Talked About
Before getting into the technical specifics, the launch method is worth examining on its own terms, because it reveals something deliberate about Alibaba’s current posture.
Qwen 3.7 Max-Preview appeared on Arena AI’s leaderboard on May 14, five days before the Cloud Summit. Alibaba sent no announcement. Developers found it, tested it, and started forming opinions. For nearly a week, Arena AI’s blind, crowd-sourced evaluation framework did the validation work:
Users comparing models without knowing which was which, voting for whichever response they preferred. Before Alibaba’s official claim, the community had already independently confirmed the model’s capabilities. Then came the marketing.
This is not how Western labs typically operate. OpenAI and Anthropic announce first, deploy later. Alibaba deployed first and let the market validate before announcing anything. The distinction matters: one approach asks for trust upfront, the other earns it before asking. It is also, as one analyst noted, marketing disguised as open research.
The substantial benchmark results, confirmed at the summit, corroborate the Arena AI data from the preview period. Qwen 3.7 Max scores 69.7% on Terminal-Bench 2.0, a test that simulates an autonomous software engineer working for five hours, 60.6% on SWE-Bench Pro based on real GitHub issues, and 92.4% on GPQA Diamond, where PhD-level physicists, chemists, and biologists write the questions. The model ranks higher than Claude Opus 4.6 and DeepSeek V4 Pro on these agentic coding benchmarks.
One claim that circulates widely and has not been independently verified: Alibaba states that Qwen 3.7 Max can operate autonomously for 35 consecutive hours and execute over 1,000 tool calls without performance degradation. That figure appears in virtually every piece of coverage written since the announcement, treated as a fact. It is not. It is an unverified marketing claim, and the AI industry’s tendency to relay vendor specifications as confirmed data deserves its own separate conversation.
The Pivot Nobody Wanted to Notice
The model’s benchmark performance is notable. The shift in Alibaba’s open-source strategy is more significant.
For three years, Alibaba played the role of the industry’s most generous open-source contributor. The Qwen model family, licensed under Apache 2.0, allowed for unrestricted downloading, modification, and deployment on any server, with no requirement to send any data to Alibaba. That strategy worked beyond anything they could have reasonably expected. The Qwen family accumulated over one billion downloads on Hugging Face, surpassing Meta’s Llama family, becoming the most downloaded AI model family in the world. Developers loved it. Startups built products on it. Several governments adopted it as a credible alternative to American-controlled platforms. Joseph Tsai, Alibaba’s co-founder and executive vice-chairman, declared in Dubai in February 2026 that open source “allows nations to claim sovereignty over AI.”
That was three months ago.
Qwen 3.7 Max is a closed, proprietary model.
No weights published.
No local deployment.
No version you can run on your own servers.
If you want to use it, you go through Alibaba Cloud’s API. This is not the first time Alibaba has done this quietly: the shift began in September 2025, when Qwen 3 Max became the first flagship to ship without open weights. Qwen 3.6 Max followed the same path. Now Qwen 3.7 Max.
The pattern is clean. This ecosystem benefits from mid-tier models remaining open, which also promotes developer adoption. The flagship is locked to generate revenue. On developer forums, the frustration is direct. One comment that circulated widely: “Qwen 3.6 Max made the entire local ecosystem better. If the Max tier stays API-only, this is a door being closed.” Alibaba has not responded.
There is a useful historical parallel. We’ve seen this before:
A tech firm provides a potent free offering, builds a dominant ecosystem on this generosity, and subsequently restricts premium access once a dependency has been created. Microsoft did exactly this with Internet Explorer in the 1990s. The difference here is that the product given away for free wasn’t a browser. It was a language model on which entire companies built their core infrastructure.
The Chip That Changes the Equation
On the same day as the Qwen 3.7 Max announcement, Alibaba’s semiconductor subsidiary T-Head unveiled the Zhenwu M890, its highest-specification AI accelerator to date.
The technical specifications are meaningful. The M890 carries 144GB of HBM3 memory and 800GB/s of inter-chip bandwidth. It is designed for both training and inference on a single chip, a capability its predecessor lacked. T-Head positions the chip against Nvidia’s H100 and H20 generations, which are exactly the parts American export controls have blocked Chinese companies from purchasing. The chip delivers roughly three times the performance of its predecessor, the Zhenwu 810E. Independent analysts note that it still trails Nvidia’s Blackwell generation, but that comparison is beside the point:
The relevant benchmark is whether it can replace what Chinese companies are no longer legally allowed to buy, and by that measure, it is functional.
More telling than the chip’s specifications is the deployment scale. T-Head has shipped over 560,000 Zhenwu units to more than 400 customers across 20 industries. This is not a prototype. It is a product already at scale.
And Alibaba published its roadmap. The V900 arrives in Q3 2027, targeting a further threefold performance increase with 216GB of memory. The J900 follows in Q3 2028. A chip roadmap three generations deep, publicly committed to, suggests the industrial seriousness that companies only announce when they are confident in their manufacturing capacity.
The political context matters here. American export controls on advanced AI chips were explicitly designed to prevent China from building a competitive AI infrastructure. The response has been to spend years developing domestic alternatives at scale. The controls may have delayed the timeline. They did not change the destination.
The 300 Million Users Alibaba Didn’t Lead With
Ten days before the Cloud Summit, Alibaba made a quieter announcement that received far less coverage in the Western press. The company connected Qwen directly to Taobao and Tmall, its consumer e-commerce platforms with over four billion listed products.
The integration is more substantive than a search upgrade. Users can ask the AI to find a product, compare prices across vendors, apply the best available discounts, place the order, and track delivery. This only action a human needs to take is payment confirmation. The agent handles the rest. The combined monthly user base of these platforms is approximately 300 million people.
This is what a full-stack AI deployment actually looks like at consumer scale: a frontier model, connected to a payments infrastructure, plugged into the largest e-commerce catalog in the world, generating behavioral data from 300 million active users that feeds back into model training. The loop is closed, and it generates revenue, data, and model improvement simultaneously.
The Stack Nobody in the West Has Built
Step back far enough from the individual announcements, and the picture that emerges is not about a single model or a single chip. It is about an integrated AI infrastructure that no longer depends on American technology at any critical layer.
Proprietary chips train proprietary models. These models run on Alibaba’s cloud infrastructure, growing at 38% annually. The API monetizes access while keeping the weights protected. These agents connect to a marketplace with four billion products and 300 million users. The user data improves the models. The models get better, which attracts more users.
The M890 chip was timed to land just before Nvidia’s quarterly earnings, a signal that this is a deliberate competitive positioning rather than an accidental coincidence of timing. Alibaba is not building toward independence from American infrastructure. In several profound ways, it has already accomplished what was intended.
A caveat that deserves acknowledgment: TSMC remains the most advanced chip fabrication facility in the world, and analysts note that the M890’s manufacturing process still relies on fabs that China does not fully control. The direction of travel is toward self-sufficiency, but the trajectory is incomplete.
The European Question Nobody Wants to Answer
I want to address something directly. When you survey what Alibaba announced across a seventeen-day window in May 2026, the natural question for anyone outside the US-China axis is: what is our position in this?
In Europe, the honest answer is uncomfortable. The strongest European AI presence is Mistral, a genuinely impressive company doing proper work with approximately 600 employees. They are trying to advance AI infrastructure in France and have announced plans for data centers and expanded compute. That is the right direction. But Mistral is a 600-person startup in a regulatory environment with a 50-billion-euro infrastructure gap, facing conglomerates that are committing fifty billion dollars over three years with near-monopoly access to favorable tax structures that American states have offered to attract data center investment.
The EU AI Act is a serious piece of regulation, and the GDPR represents a genuine and necessary set of protections. But regulation alone does not train a model. Europe currently has no frontier AI chip, no hyperscale cloud infrastructure of comparable size, and no consumer platform with 300 million users. What the continent has, to be direct about it, is exceptional skill at regulating technologies built elsewhere.
The energy dependency parallel is worth taking seriously. For decades, Germany built its industrial competitiveness on Russian gas: cheap, reliable, abundant, and controlled by a country that would eventually become an adversary. Their response, when the dependency became visible overnight, was to replace Russian gas with American LNG and Gulf petroleum. The dependency was not solved. The supplier was changed.
Europe is repeating this pattern with AI infrastructure, transitioning from exclusive American dependency toward a menu of American or Chinese options, neither of which represents actual sovereignty.
What Alibaba has demonstrated with Qwen 3.7 is not only that a Chinese model can compete with the best American models. DeepSeek established that in February 2025. What Alibaba has demonstrated is that it is possible to build the entire AI value chain from semiconductors to consumers without passing through Silicon Valley and to do it at roughly half the price. That is the genuinely new information from these announcements.
What This Means for Anyone Using These Tools Today
The Qwen 3.7 Max API is available today from Europe at $2.50 per million input tokens and $7.50 per million output tokens. A workflow that costs $300 in Claude Opus tokens runs for approximately $50 on Qwen 3.7 Max. That is a six-to-one cost ratio for comparable or, in some agentic coding benchmarks, superior performance.
For a startup that continuously runs AI agents, that ratio changes the economics of the entire product. It is not a question of model preference. It is a question of whether the business can survive at one pricing structure but not the other.
The open-weight Qwen 3.6 models remain available for self-hosted deployment in Europe, with no data routing through either American or Chinese infrastructure. That is a solid argument for companies with strict data residency requirements. Behind Alibaba Cloud’s API is the flagship model, and any data that travels through it might fall under Chinese data regulations.
The tools are more powerful, more accessible, and cheaper than at any prior point. The sovereignty question is not whether to use them. It is under whose infrastructure you choose to run them, and whether that choice is made deliberately or by default.
That choice, made at scale across European enterprises over the next two years, will determine whether the continent’s digital infrastructure ends up in a more or less dependent position than it occupies today.
Where do you think Europe’s realistic position in this race actually is? I’m curious whether you see the regulatory advantage as a constraint or a long-term competitive asset.