Qwen

Alibaba's Qwen is the world's broadest open-weightA model where the trained weights are freely downloadable — you can run it yourself without contacting the creator. Llama, Mistral, Qwen, and Gemma are open-weight. Open-weight does not mean open-source: the training data and code often stay private. The license still governs what you can do with the weights, including whether you can use them commercially. model family — Apache 2.0 weightsThe numerical values inside a trained model that encode everything it has learned. A model is, functionally, a giant list of weights — tens of billions of numbers for a mid-sized model, hundreds of billions for a frontier model. "Open-weight" means those numbers are published. "Downloading the weights" means getting the actual file you'd need to run the model yourself. spanning phone-size to 397 billion parameters — now paired with a closed, API-only frontier flagship.

Qwen (from the Chinese "Tongyi Qianwen") is the large-model team inside Alibaba Cloud, the cloud-computing arm of Alibaba Group. Since 2023 it has been the most prolific frontier-model shipper of any major tech company, releasing models at a pace that outstrips most dedicated AI labs — dense models, mixture-of-experts models, coding specialists, vision-language models, audio models, and embeddings, across more than a dozen size points.

The strategy behind that firehose is straightforward and worth understanding, because it explains why the models are so generously licensed. Alibaba Cloud makes its money from cloud compute and API access, not from selling model licenses. Open-sourcing Qwen drives adoption; adoption drives people to run those models — often on Alibaba Cloud. The result is that Qwen has become one of the most-downloaded model families on Hugging Face and the default open-weightA model where the trained weights are freely downloadable — you can run it yourself without contacting the creator. Llama, Mistral, Qwen, and Gemma are open-weight. Open-weight does not mean open-source: the training data and code often stay private. The license still governs what you can do with the weights, including whether you can use them commercially. choice for a huge swath of developers and businesses worldwide.

For most of its history, Qwen's answer to "how open are you?" was "very": nearly the entire lineup ships under Apache 2.0, the gold standard of permissive licenses — unrestricted commercial use, modification, and redistribution, no royalties, no user-count carve-outs. That is more permissive than Meta's Llama license and on par with the most open models anywhere.

In 2026 the posture got more nuanced. Alibaba began holding its absolute frontier models closed: the agent-tuned "Max" flagships (Qwen3.6 Max, then Qwen3.7-Max) and their multimodalA model that can handle more than one type of input or output — typically text plus images, sometimes plus audio or video. "GPT-4 Vision" and "Llama 3.2 11B Vision" are multimodal models that accept both text and images. A text-only model is called "unimodal" but nobody uses that term; text-only is the assumed default. "Plus" siblings are proprietary and API-only, with no downloadable weightsThe numerical values inside a trained model that encode everything it has learned. A model is, functionally, a giant list of weights — tens of billions of numbers for a mid-sized model, hundreds of billions for a frontier model. "Open-weight" means those numbers are published. "Downloading the weights" means getting the actual file you'd need to run the model yourself.. The tier just below — the numbered open-weightA model where the trained weights are freely downloadable — you can run it yourself without contacting the creator. Llama, Mistral, Qwen, and Gemma are open-weight. Open-weight does not mean open-source: the training data and code often stay private. The license still governs what you can do with the weights, including whether you can use them commercially. models like Qwen3.5-397B-A17B and the Qwen3 and Qwen3.6 families — stays Apache 2.0. So the lineup now splits cleanly: open-weight workhorses you can self-host freely, and a closed frontier model you can only rent through the API. For a business reader, that split is the single most important thing to keep straight.

Match the model to the job, and mind the open/closed line. If you want maximum capability and you're comfortable using a hosted APIAccessing a model by sending requests to the creator's (or a provider's) servers, typically pay-per-use. Hosted APIs handle all the operational work — scaling, hardware, uptime — in exchange for a per-token or per-request fee. Every closed-API model is hosted; many open-weight models are also available via hosted APIs from providers like Together, Fireworks, or Groq., the closed Max flagship is the top of the range. If you want to own your stack — self-host, keep data in-house, avoid vendor lock-in, fine-tuneA model that has been further trained on additional data to specialize it for a particular task, domain, or style. Fine-tuning a general model on medical literature produces a medical specialist; fine-tuning on your company's support tickets produces a support assistant that sounds like your team. Fine-tunes are much cheaper to create than training a model from scratch. freely — the open Apache 2.0 models are the reason Qwen is so widely used, and they run on everything from a laptop to an H100 cluster.

The China-jurisdiction consideration applies the same way it does for any Chinese lab: the hosted DashScope API routes data to Alibaba Cloud under Chinese law, while the open weightsThe numerical values inside a trained model that encode everything it has learned. A model is, functionally, a giant list of weights — tens of billions of numbers for a mid-sized model, hundreds of billions for a frontier model. "Open-weight" means those numbers are published. "Downloading the weights" means getting the actual file you'd need to run the model yourself., run on your own infrastructure, carry no such routing. Qwen has drawn far less government-restriction attentionThe mechanism inside a Transformer that lets the model weigh which parts of the input matter most when processing each word. When you read "the cat sat on the mat," attention is how the model knows that "it" in a later sentence refers back to the cat, not the mat. Attention is what made modern language models possible. than DeepSeek did, but the underlying data-governance logic is identical — and for the closed Max/Plus models, the hosted API is the only way to use them, so there's no self-host escape hatch for those specific models.

Against Meta, Qwen is more permissively licensed across its open tier (clean Apache 2.0 versus Llama's community license with its large-user carve-out) and offers a far wider range of sizes, but carries the China-jurisdiction consideration Meta doesn't. Against DeepSeek, the two are the leading Chinese open-weightA model where the trained weights are freely downloadable — you can run it yourself without contacting the creator. Llama, Mistral, Qwen, and Gemma are open-weight. Open-weight does not mean open-source: the training data and code often stay private. The license still governs what you can do with the weights, including whether you can use them commercially. labs — DeepSeek tends to win on raw frontier capability-per-dollar and keeps its flagships MIT-open, while Qwen wins on breadth of sizes, Apache licensing, and multilingual coverage, but has moved its very top models closed. Against the Western closed labs (OpenAI, Anthropic, Google) and Mistral AI, Qwen's open tier is the pitch: competitive capability you can download and self-host for free, in exchange for the data-governance questions a US- or EU-based vendor doesn't raise.