What Is Qwen? Choosing Hosted APIs or Local Models
Qwen is a model family from Alibaba Cloud’s Qwen team, with both hosted services and open-weight releases. Using Qwen3, this article explains model naming, thinking controls and local deployment choices.
Meet this model
Qwen is the English name of the Tongyi Qianwen model family. It does not identify a single model size and is distinct from the Qwen chat app. Cloud platforms provide callable model services, while official repositories publish weights for particular releases. For Chinese document extraction or coding assistance, choose a specific model and deployment method. Hosted names, open-weight parameter sizes and app features are not interchangeable.
Choose hosted models and open weights separately
Alibaba Cloud Model Studio lists general models alongside specialized media models. Its October 20, 2026 catalog includes hosted names such as qwen3.8-max, which does not by itself establish that weights with the same name are available for download. For offline use, find the full repository name under the official Qwen organization and inspect its input types and model card instead of relying on a marketing name.
The official Qwen3 repository is a verifiable open-weight example, offering multiple model sizes and local running instructions. It states that its open-weight releases use Apache 2.0. That statement applies to the listed releases, not automatically to every historical Qwen model or future hosted model. Retain the license accompanying the model you download.
Understanding thinking controls through Qwen3
The Qwen3 quickstart provides an enable_thinking control in the chat template. When enabled, the model generates thinking content before the final response; disabling it selects non-thinking behavior. This differs from asking “do not explain” in a user message: one changes the invocation configuration, while the other requests a presentation style.
For example, when classifying support messages into business categories, request one allowed label and compare errors and response time with thinking enabled and disabled. For mathematical derivations, check intermediate conditions and the final result. Do not copy Qwen3 controls to every Qwen variant. Separate Thinking or Instruct releases require their own calling instructions.
Local use also depends on templates and memory
The documentation covers execution through Transformers and local chat tools. Loading weights also requires the matching tokenizer and chat template. Simply concatenating user text can omit the role markers expected by the model. That formatting error should not be casually attributed to weak Chinese-language ability.
Smaller or quantized models may be easier to try on constrained hardware, but precision and context length affect the running conditions. Check extraction on a passage containing Chinese identifiers and amounts before deciding whether it suits the task. This article does not download or run these models, nor claim that an ordinary computer will reproduce official benchmark results.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.