Alibaba Has Open-Sourced More Than 460 Models and Spawned 300,000 Derivatives

Alibaba has open-sourced more than 460 AI models. Those models have, the company says, spawned more than 300,000 derivatives — fine-tunes, quantizations, merges and repackaged builds made overwhelmingly by developers who do not work for Alibaba and do not live in China. It is the largest deliberate giveaway of near-frontier AI capability anyone has attempted, and it has quietly turned a Hangzhou e-commerce conglomerate into the default substrate of the global open model ecosystem.

Both figures come from Alibaba itself, released in an emailed statement on Aug. 15 alongside the headline claim that its open-weight Qwen family has passed 3 billion cumulative downloads, overtaking Meta, Alphabet and every domestic Chinese rival to become the world’s most-used AI model family. The claim is directionally correct and specifically inflated, and the gap between those two things is the most interesting part of the story.

Hugging Face published its State of Open Models report on Aug. 14, one day earlier. It counts 2,045 million Qwen downloads on its Hub in 2026 — 2,061 million including every associated repository. Google’s models drew roughly 418 million. Meta’s drew roughly 227 million. Qwen is not narrowly ahead. It is ahead of Google by nearly five to one and of Meta by nearly nine to one.

But 2.045 billion is not 3 billion. Alibaba’s number spans every platform, including its own ModelScope hub, which Hugging Face cannot see; Hugging Face’s count deliberately excludes API traffic, private enterprise deployments and models distributed elsewhere. Both numbers are defensible. Only one made it into the press release, and it is about a third larger than the platform being cited actually recorded.

The same asymmetry runs through the derivative count. Against Alibaba’s 300,000-plus, Hugging Face counts 151,448 Qwen-derived models on its Hub. Even the smaller number is remarkable: it is 2.6 times Meta’s entire open-source footprint on the platform, 4.7 times the number of Llama-specific repositories, and nearly double Google’s 82,506 derivatives. The report’s own verdict is that Qwen constitutes “one of the largest foundations of the open AI ecosystem,” and that “Qwen has become part of the default workflow for developers deciding what models to fine-tune and deploy.”

The report also explains the mechanism. Of 178 Chinese open models above 20 billion parameters that Hugging Face tracked, 59% shipped under Apache 2.0 and 22% under MIT. None carried a restriction limiting use to non-commercial purposes. Meanwhile the largest open models released by leading Chinese labs ranged from 754 billion to 2.78 trillion parameters in a given month, while the biggest American open releases stayed below 130 billion for most of the year. Chinese labs are shipping larger models under looser licences — and Alibaba pushes its through Alibaba Cloud into enterprise customers across Southeast Asia and Africa, a distribution channel most rivals simply do not have.

What a derivative actually measures

It is worth being precise about what 300,000 derivatives is and is not. It is not 300,000 companies building products on Qwen. The overwhelming majority of derivative repositories are quantized builds, LoRA adapters, format conversions and community repackages — the work of an intermediary layer of groups like lmstudio-community and mlx-community that compress and re-host models for specific hardware. A single upstream release can generate dozens of downstream artifacts within days.

That does not make the number meaningless. It makes it a measure of a different thing: not revenue, but gravity. Derivative counts track which base model a developer reaches for by reflex when they need something to fine-tune, and reflexes are extremely sticky. Every quantized Qwen build that lands in a local-inference toolchain is a small piece of infrastructure that makes the next Qwen release easier to adopt than the alternative. Downloads measure attention. Derivatives measure lock-in.

This is precisely why the US–China open-weight question has stopped being about benchmarks. Hugging Face chief executive Clément Delangue put the position bluntly to CNBC on Aug. 3: “They’re clearly dominating on open models right now, and I wouldn’t be surprised if they start dominating at the frontier either by the end of this year or next year at the rate of progress.” Hugging Face now attributes roughly 41% of open-weight downloads to Chinese developers. American companies route more than 30% of their OpenRouter tokens through Chinese open-weight models, up from 4.5% in the first half of 2025. The substrate question has already been answered in a lot of production stacks.

American labs have noticed. Meta and Nvidia have both shipped new open models in recent weeks, with Nvidia explicit that it is chasing China on open weights. That is a reactive posture, and it arrives after the ecosystem has already formed around someone else’s file formats.

What to watch

The sharpest threat to Qwen’s run is not Meta. It is Alibaba.

The company that built this ecosystem has spent 2026 walking away from the strategy that produced it. Flagship models including Qwen3.7-Max and Qwen3.7-Plus went closed and API-only. In mid-March, Alibaba folded Tongyi Lab, Qwen and three other units into a new Alibaba Token Hub group reporting to chief executive Eddie Wu, whose internal mandate, per a memo reviewed by Bloomberg, was to “create tokens, deliver tokens, and apply tokens.” Alibaba Cloud chief technology officer Zhou Jingren was installed over the AI unit with a commercial brief. Qwen technical lead Junyang Lin left on March 3 with a six-word post: “me stepping down. bye my beloved qwen.” Alibaba has since floated revenue-sharing terms and charging its heaviest open-model users.

The second threat is Beijing, which has been weighing curbs on overseas access to China’s best models — restrictions that would remove exactly the international reach that produced these numbers.

So the 460 models and 300,000 derivatives are best read as a completed campaign rather than an ongoing one. Alibaba won the open substrate. The open question for the rest of 2026 is what it charges for having won it, and whether a developer base assembled on permissive licences stays put when the terms change.