Qwen models: names, capabilities and selection

Qwen is more than one chatbot model

The official catalog groups models by task. Language generation, visual understanding, image generation and audio generation are different capabilities. For example, it lists qwen-image-3.0-pro for image generation and qwen-audio-3.0-tts-plus for text-to-speech. Those are not substitutes for the language-model ID in a basic text chat request.

A model that reads a screenshot does not automatically generate pictures. A speech-generation model does not automatically transcribe recordings. Start with the input and output your application actually needs, then open the corresponding model reference.

The current language guide places several earlier families, including Qwen3-Coder variants, in its legacy section. Legacy is not the same as discontinued. An existing deployment may still use one; consult the provider’s status and migration notices before replacing it. A specialized name alone is not evidence that it will outperform every newer general model on your coding task.

Hosted API names and open weights are different

Downloadable Qwen releases use repository names such as Qwen/Qwen3-8B and Qwen/Qwen3-30B-A3B-Instruct-2507. These are established examples from the official Qwen3 repository, not a claim that they are the newest releases.

Model names can describe a size, an architecture, a task or a revision. In Qwen3’s mixture-of-experts naming, 30B-A3B distinguishes total parameters from the activated subset. Do not interpret the active count as the complete memory requirement. Weights, precision, context cache and serving overhead all affect hardware needs.

For a download, record the exact repository and revision, check that model’s license and model card, and confirm support in your chosen runtime. A quantized file or a third-party runtime tag may not have the same name or defaults as the original release.

A practical way to choose

1. Choose the deployment route. Use a managed API when you want the provider to run inference. Consider local or self-hosted weights when you need control over the runtime and can support the hardware and operations.
2. Define the task and failure condition. Examples include returning valid fields from an invoice, fixing a bug while preserving tests, or answering a question using only supplied material.
3. Compare a small shortlist. For a hosted text workflow, evaluate a Flash option, Plus and Max on identical inputs. Their names are starting points for testing, not an independent ranking.
4. Measure the whole request. Record correctness, response time, input and output tokens, retry frequency and total cost. Keep the region, model ID and thinking settings fixed during a comparison.
5. Retest before switching. Keep difficult examples as a regression set so a version change does not quietly break a working workflow.

These are our suggested evaluation steps. ChinaAI has not published head-to-head benchmark scores for this shortlist. See how we distinguish specifications from testing.

What does a 1M context window mean?

It describes available token capacity, not one million words, guaranteed perfect recall, or an answer of that length. Requests include instructions and history as well as your main document. Output limits and thinking budgets need separate attention.

Begin with the smallest useful excerpt and a defined output format. For repeated questions over a large collection, evaluate retrieval instead of repeatedly sending every document. Check the selected model’s limits and the price tier triggered by your request length.