All Alibaba Cloud (Qwen) models available through the API

Choose a Model Studio candidate by its task, endpoint and exact identifier. Keep aliases, snapshots and hosted-model relationships distinct when evaluating a Qwen integration.

Model families at Alibaba

The recommended catalog includes Qwen general-purpose models, visual and audio capabilities, generation models, embeddings and reranking, alongside hosted third-party entries. Official documentation.

Start with the input and output required by the product. A model that understands an image is not automatically an image-generation endpoint. A speech-to-text stage and a conversational audio stage also have different request contracts. Keep these distinctions in the shortlist before comparing rates.

Qwen Coder documentation describes coding-oriented models and model-specific billing and caching behavior. Official documentation.

Use a coding candidate when the evaluation fixture involves a verifiable project change. Require the patch to pass relevant checks and include a case that needs the model to ask for missing context or stop. A polished explanation around an invalid patch should not count as an accepted result.

For retrieval work, evaluate embedding and reranking choices as parts of the retrieval pipeline. The final answer model cannot compensate reliably for evidence the pipeline never retrieves. Keep retrieval quality and generation quality as separate observations when comparing the application.

All models

Copy the exact identifier and inspect the deployment scope available to the workspace. A similarly named model in another service scope may not be accessible through the current key and host. Keep that access relationship with the record instead of discovering it only during production deployment.

Use the model page’s source to investigate an unknown capability. If evidence is insufficient, retain the uncertainty. A blank field should not become a positive feature claim simply because another member of the family supports it.

Cheapest / strongest / longest-context

These rankings describe recorded attributes. The cheapest input category may suit a document-heavy job but fail to capture a response-heavy workflow. The largest capacity may reduce the need to split a document without establishing that the model uses every passage reliably.

Define strongest through the application’s acceptance criteria. For extraction, verify fields and source evidence. For summaries, preserve required facts and distinguish missing material. For code, evaluate a working patch. Use the data list to select candidates, then keep the observed outcomes in a separate evaluation sheet.

Compare candidates under the same response contract. A model that emits an extensive explanation is not producing the same workload as one instructed to return a narrow structured object. Fix the requirement, inspect validity and measure actual usage before drawing a cost conclusion.

Deprecations and replacements

Model Studio publishes model update information and links deprecation guidance from relevant error and model documentation. Official documentation.

When an identifier is retired, search environment settings, scheduled jobs and client presets for every consumer. Updating the visible example alone can leave old background work unchanged. Keep the replacement’s request settings and workspace access in the migration checklist.

Retain the original safe fixture before changing the model. First verify the new route with a minimal request, then restore required structure, media and tools separately. If an output changes, compare accepted behavior rather than assuming the new model should reproduce the old wording.

A historical model page should preserve the retirement context and replacement route. It remains useful for diagnosis, but it should not be presented as an active recommendation when the provider no longer supports it.

Naming and versions

Treat dated snapshots and current aliases as explicit version choices. A snapshot supports a more controlled evaluation record, while an alias can change its target under the provider’s model lifecycle. Inspect the current catalog and release notes before relying on a naming convention alone.

Preserve offer and capacity conditions with the exact identifier. Version changes can affect more than output behavior: the account-visible free allocation and rate row may differ. Keep those operating changes in the migration record so a successful code edit does not hide a changed usage assumption.

For hosted third-party models, preserve the host as well as the underlying family. A Model Studio route and the model developer’s direct API can use different access and billing arrangements. Evaluate the endpoint you will actually operate.

Review model-specific pricing, quota scope and the workspace-based Python baseline with the selected identifier.

Use the AI API cost calculator to turn the model and workload you are considering into an estimate.

Last verified 12 Sep 2026 · Source ↗

Frequently asked questions

Sources

• First Qwen API call ↗
• Recommended models ↗
• Model inference pricing ↗
• API keys and permissions ↗
• Rate limits ↗
• Error codes ↗
• New-user free quota ↗
• Model usage ↗
• Context cache ↗
• Batch API ↗
• Billing and cost management ↗
• Dynamic rate limiting ↗
• Qwen Coder ↗
• Model updates ↗

Last verified 12 Sep 2026 · Source ↗