Tongyi
Tongyi Qwen, developed by Alibaba Cloud, is a sophisticated series of LLMs. It includes multiple variants, such as Qwen for text processing, Qwen-VL for vision-language tasks, and Qwen-Audio for audio understanding. The models are notable for their impressive scale, with the flagship Qwen-72B model featuring 72 billion parameters and trained on over 3 trillion tokens.
After installation, you need to get API keys from Alibaba Cloud and setup in Settings -> Model Provider.
JSON Schema is available in the model settings for Qwen3.7 Max/Plus/Flash (including the registered snapshots), Qwen3.8 Max/Flash, DeepSeek V4.1 Flash/1017, and DeepSeek V4 Pro.
DeepSeek V4 Flash 0731 supports JSON Object only.
The supported formats follow Alibaba Cloud's structured output guide and DeepSeek feature table.
For text input, the plugin sends the schema through the existing DashScope endpoint with inside .
Omitting defaults to ; an explicit is preserved.
The schema received by the plugin, including nested objects, arrays, optional fields and enums, is passed through unchanged.
For media inputs supported by the selected model, Alibaba Cloud only guarantees JSON syntax.
The plugin uses JSON Object and adds the complete schema as an instruction, preserving the original messages.
Field names, types and other schema constraints are not guaranteed for these inputs.
Plain text sent through the multimodal endpoint still uses JSON Schema.
Before deploying workflows that require strict constraints, verify the target model and region against the real service:
The official guide explicitly documents Beijing OpenAI-compatible Chat Completions support for DeepSeek V4.1 Flash/1017.
Offline tests verify registration, HTTP serialization and response handling; they do not establish server enforcement on the native DashScope endpoint or in other regions.
Dify 1.13.3 rewrites boolean schema fields before invoking plugins, so verify the schema sent by the Dify version used in production.
Provider credential validation now uses , which is available in
Beijing and Singapore, instead of the retired . Existing
overrides are still honored for keys restricted to another model.
The predefined catalog was checked against Alibaba Cloud's retirement notices on
October 13, 2026. Retired model definitions and their ordering entries have been
removed, including older snapshots and their legacy short-date aliases. Existing
apps using these models need to select a supported replacement; their model IDs
are not automatically rewritten.
See the current model catalog,
Qwen3.8-Flash specifications,
and Qwen-MT-Flash specifications.
Model availability depends on the API key's region and permissions.
Tongyi speech-to-text patches DashScope's async-to-sync websocket bridge when it runs under Dify's gevent-based plugin runtime. This avoids sustained high CPU after STT requests while keeping the normal in-process recognition path.
If you need stronger isolation for STT recognition, set in the plugin daemon environment.
can be used to tune the subprocess timeout.
is optional and disabled by default. Turning it on attaches and to each Tongyi request as custom HTTP headers, so usage can be attributed to a specific Dify app in the DashScope server logs. The opt-in is composed with the existing bury-point header (which carries market analytics) in a single merged headers dict before the call to or . The Dify session lookup is best-effort: if the session context is not initialized, no Dify headers are attached and only the bury-point header is sent.
Some outbound calls build their address at runtime, so the list may be incomplete.
Last checked 11/18/2026 09:55 AM · version 0.2.29