Qwen-Long supports 10 million tokens. Alibaba’s four file routes work differently

“Qwen supports files” can refer to several products with different upload limits, retention rules, model behaviour and API contracts. A browser File Q&A session, a generic agent File API, Qwen-Doc-Turbo and Qwen-Long are not four interfaces around one storage system. They are separate routes designed for different tasks.[2][3][4][5]

This fragmentation matters because a working prototype can fail when moved into production. A file that survives indefinitely in Qwen-Long may disappear after 30 days in the general File API or immediately when a File Q&A page is refreshed. A 150 MB document accepted by Qwen-Doc may not fit the 20 MB direct-upload limit of the general agent file endpoint.

A file limit is not a context limit

Qwen-Doc-Turbo allows a 150 MB file but caps document input at 262,144 tokens. Qwen-Long accepts up to 10 million tokens through file references and bills the referenced content on each call. The Qwen-Long documentation used here applies specifically to mainland China’s Beijing region and requires a key for that region. Neither its capacity nor its data location should be assumed for another Alibaba endpoint.[3][5]

Qwen-Long’s upload mechanism avoids retransmitting large bytes with every request, but its documentation is explicit that file content tokens are counted as input on every call. Storage convenience and token cost are different properties. A repeated ten-million-token query does not become cheap because the file ID is short.[5]

OpenAI-compatible still means Qwen-specific behaviour

Alibaba Cloud’s Responses documentation states that only parameters explicitly listed on the page are processed and other OpenAI parameters are ignored. The route is synchronous and does not support the background parameter. That reduces migration work at the client-library layer while requiring the application to validate which controls actually exist.[1]

The same API supports Qwen-specific built-in tools, previous_response_id, stored responses and session caching. An application cannot infer those behaviours from generic OpenAI examples or from Qwen-Long’s older file-reference flow. The provider’s current route, region and model list are part of the contract.

The product name should appear in the architecture diagram

A durable implementation should record the exact upload endpoint, region, purpose, parser, model, retention policy and deletion path. “Stored in Qwen” is not enough for a security review or a recovery plan. The same filename can exist in a temporary browser session, a 30-day agent workspace and a long-context account store with different identifiers and deletion rules.

Evaluation should also use the selected route. A browser File Q&A demonstration cannot validate Qwen-Long’s ten-million-token retrieval, and a Qwen-Doc summary cannot establish that an agent’s file-search tool returns the same passages. Each route needs known-answer tests that match its parsing and retrieval method.

Qwen’s breadth is a genuine advantage: there are routes for quick browser questions, agent files, document analysis and ultra-long context. The limitation is that the umbrella name conceals separate systems. Reliability improves when the product boundary is made explicit rather than hidden behind one claim that the model “supports files”.

Report an error or suggest a correction