--- title: "Alibaba gives Qwen3 to the world, but keeps its largest model behind a paywall" date: 2026-11-08 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-alibaba-qwen-opensource language: en --- # Alibaba gives Qwen3 to the world, but keeps its largest model behind a paywall > Alibaba open-sourced the Qwen3 family under Apache 2.0, yet reserved its trillion-parameter Qwen3-Max for paid APIs — a split that shapes who can actually build on Chinese open-source AI. The spring of 2025 felt like a dam breaking for open artificial intelligence in China. A small research team could suddenly download a frontier-grade model, run it on a laptop-class GPU, and call it their own. The catch arrived quietly, buried in the same announcement that gave so much away for free. ## What happened On 28 April 2025, Alibaba's Tongyi (通义) lab released Qwen3, the third generation of its large model (大模型) family, and made the core of it freely available under the Apache 2.0 license. Unlike many "open" releases that ship only a chat interface or a restricted weight file, Alibaba published both the weights and the code for eight models at once. The two headline models are mixture-of-experts (MoE) systems. The flagship Qwen3-235B-A22B carries 235 billion total parameters but activates only 22 billion per token, which keeps inference cheap relative to its size. A smaller sibling, Qwen3-30B-A3B, holds 30 billion parameters and activates 3 billion. Alongside them, Alibaba open-weighted six dense models — 32B, 14B, 8B, 4B, 1.7B and 0.6B — all under the same permissive license. ## Why the license matters Apache 2.0 is the detail that separates a gift from a teaser. It permits commercial use, modification and redistribution without per-seat fees, and it includes an express patent grant. For a startup in São Paulo or a hospital in Riyadh, that means the model can enter a product without a licensing negotiation. By late April 2025, Alibaba said the broader Qwen line had already produced more than 200 open-source (开源) models, passed 300 million global downloads and spawned over 100,000 derivative models — a footprint the company describes as the largest open-source model family in the world. Those numbers come from Alibaba's own FY2025 earnings disclosure, not an independent count, but they track a visible trend on hosting platforms such as Hugging Face and ModelScope. ## The twist: not everything is free Here is the part easy to miss. Qwen3-Max, the proprietary sibling with more than one trillion parameters, was held back from the open release and offered only through Alibaba's paid API. The open models are powerful; the largest, most capable variant is a cloud product. Alibaba is effectively running a two-track strategy: give the community enough to build, train and compete, while reserving the top of the stack to pull usage onto its cloud computing (云计算) platform. ## Who actually benefits The 119-language coverage and 128K context window — present on all but the three smallest models — matter because they let non-English teams build without first translating their whole stack. A logistics firm in Indonesia can fine-tune Qwen3-14B on local address formats. A clinic in Brazil can keep patient notes in Portuguese entirely on-premise. The point is not just capability; it is permission to operate without asking a vendor first. ## Hybrid thinking Qwen3 introduced a switch between a "thinking" mode for hard reasoning and a "non-thinking" mode for fast replies, controllable per request. This lets developers trade latency and cost against depth — a practical lever when serving millions of calls, and a quiet admission that one model size does not fit every bill. ## Key takeaways - Alibaba open-sourced eight Qwen3 models under Apache 2.0 on 28 April 2025, including MoE flagships and six dense sizes. - The free tier covers models up to 235B total parameters (22B activated); the trillion-parameter Qwen3-Max stays API-only. - Open-weight releases are a distribution channel: they seed adoption that Alibaba monetizes through cloud usage. - Alibaba Cloud booked 118 billion RMB (about 16.4 billion US$ roughly 128 billion HKD) in FY2025 revenue, up 11% year on year, with AI-product revenue growing triple digits for seven straight quarters. - Alibaba has pledged 380 billion RMB (about 53 billion US$ roughly 413 billion HKD) over three years for cloud and AI infrastructure. ## What readers can do now - **Prototype locally first.** Download Qwen3-8B or Qwen3-14B and run a private pilot before committing budget to any vendor API. - **Budget for the cloud gap.** If you need Qwen3-Max-class capability, price the API and compare it against self-hosting the open MoE before signing. - **Check the license fine print.** Apache 2.0 is permissive, but confirm export and sector restrictions for your jurisdiction before shipping. - **Mirror the weights you depend on.** Copy Qwen3-8B or Qwen3-14B into your own registry on Hugging Face or ModelScope so a future policy change cannot yank a pinned version out from under a running service. - **Toggle thinking mode per request.** In a prototype, switch the thinking and non-thinking modes on the same prompt and log the latency and token-cost delta, so you ship the cheaper mode by default and reserve deep reasoning for the calls that actually need it. ## Why it matters For the ordinary reader, the practical upshot is choice. Because the weights and code are public, no single company can quietly rewrite the terms of a service you already rely on; you can freeze a version, inspect it, or move to a fork. For a developer, the gap between the open mixture-of-experts models and the API-only flagship is a design constraint rather than a disappointment: it shows exactly where to spend your own engineering effort. Build the bulk of a product on the open weights, learn the shape of the family, and reach for the paid tier only when a specific task truly needs the largest variant. That habit turns a corporate release into a reusable skill instead of a recurring bill — and it is why the open tier, not the paywall, is the part worth studying first. ## Honest limitations This article relies on Alibaba's official release notes and its FY2025 earnings figures; the "largest open-source family" claim is the company's own framing. Benchmark comparisons between Qwen3 and rival models shift constantly, and we have not independently reproduced them. The cloud-revenue and infrastructure-investment figures are denominated in RMB and converted here at approximate rates; treat them as directional, not audited cross-currency equivalents. We did not have access to per-model training cost or energy data. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-alibaba-qwen-opensource Free to quote with attribution and a link to the original.