Qwen
AIBring Alibaba's Qwen models into your flows via Model Studio. Agents generate text and run function-calling loops, analyze images with Qwen-VL, create embeddings, produce Wan images and video, and convert speech both ways with TTS and ASR.
What This Integration Enables
Qwen is the connector for teams whose customers do not all speak English. Alibaba's Model Studio puts an unusually complete multimodal roster behind one API key: the Qwen chat family for text and function calling, Qwen-VL for vision, Qwen3 embeddings, Wan models for image and video, and speech in both directions, with TTS that covers Chinese regional dialects and ASR that detects language and emotion automatically. For a company operating in APAC markets, that is not a nice-to-have. It is the difference between localizing content and actually producing it. FlowRunner treats Qwen the way it treats every model provider under its BYOK approach: as a capability inside an orchestrated flow, not as the flow itself. The model generates; the orchestration decides what happens next, logs what was decided, and routes to a person when the output is headed somewhere public. - Run single-prompt or fully custom multi-turn chat, including tool calling and thinking mode with a token budget - Analyze images, OCR documents, and reason over charts with Qwen-VL models - Generate embeddings for semantic search with configurable dimensions up to 2048 - Render images synchronously or submit image and video tasks asynchronously and poll them - Convert text to multilingual speech and transcribe audio with biasing context and language hints
Without FlowRunner
With FlowRunner
Use Case Scenarios
The voicemail queue that sorts itself
Call recordings from the support line arrive in a shared mailbox. An agent runs Transcribe Audio on each one; qwen3-asr-flash detects the spoken language automatically and reports detected emotion alongside the text. Transcripts land in [Google Sheets](/integrations/google-sheets) with language, emotion, and a Chat Completion summary attached. Recordings flagged with frustrated or angry emotion post to [Slack](/integrations/slack) immediately, where a support lead decides who calls back first. The queue that used to be listened to in order is now read in priority order.
Document intake for a bilingual back office
Supplier invoices and shipping documents arrive as photos and scans, some in English, some in Chinese. Analyze Image extracts the text and structure with qwen-vl-max, Chat Completion in JSON mode normalizes the fields, and inverse text normalization keeps amounts and dates formatted correctly in both languages. Each document's text also flows through Create Embeddings into a vector store such as [Qdrant](/integrations/qdrant), so the next agent that needs to answer "have we seen this supplier before" searches meaning, not filenames.
The script-to-video content line
Chat Completion drafts a 30-second product script from a feature brief. Create Video Task submits the render to a Wan model and the flow moves on; Get Task Status polls until the status reads SUCCEEDED, typically within a few minutes. Because DashScope result URLs expire after 24 hours, the finished video is immediately archived to [Dropbox](/integrations/dropbox-service) with the script and seed recorded alongside it, so a good result can be regenerated deliberately rather than remembered wrongly.
Human-in-Loop Highlight
Synthesize Speech will happily produce a confident Cantonese voiceover for a team where nobody speaks Cantonese, and Create Video Task will spend real money against the Model Studio account on every render, whether the script deserved it or not. So the content line gates twice. Before a batch of video renders is submitted, a person approves the scripts, because each task is spend that cannot be recalled once submitted. And before any generated voiceover or video reaches a customer-facing channel, the flow routes the asset to a reviewer who actually speaks the target language, with the source copy side by side. The models produce at machine speed. The decision that a piece of generated media represents the company in a language the team cannot verify belongs to a person who can.
Agent Capabilities
11 actionsChat
2- Chat Completion Generates a text response for a single prompt using a Qwen model (qwen-max, qwen-plus, qwen-flash, qwen-turbo, qwen3-max and others). Supports an optional system prompt, thinking mode with a token budget, JSON mode for structured output, sampling controls, and stop sequences. Returns the text, reasoning content when thinking is active, finish reason, and token usage.
- Chat Completion (Advanced) Sends a fully custom request with a complete messages array: multi-turn conversations with system, user, assistant, and tool roles, multimodal content parts for Qwen-VL, tool and function calling passthrough, a structured response format object, and thinking controls. Returns the raw API response including tool calls. The action behind agent loops.
Vision
1- Analyze Image Analyzes one or more images with a Qwen vision-language model and answers a text prompt about them: describe scenes, extract text with OCR, compare images, identify objects, or reason over charts and documents. Images are passed as public URLs or base64 data URLs.
Embeddings
1- Create Embeddings Generates vector embeddings for up to 10 texts using text-embedding-v4 (Qwen3-Embedding) or text-embedding-v3, with configurable output dimensions from 64 to 2048 depending on model. Returns vectors in OpenAI format with token usage.
Image Generation
2- Generate Image Generates images from a text prompt using a Wan text-to-image model and waits for completion. Because DashScope result URLs expire after 24 hours, generated images are automatically saved to FlowRunner file storage with permanent URLs. Supports negative prompts, custom sizes, 1 to 4 images per request, prompt extension, and seed control.
- Create Image Task Submits an asynchronous text-to-image task and returns immediately with a task ID for non-blocking flows or custom polling logic. Task IDs and result URLs remain valid for 24 hours, so results should be downloaded promptly.
Video Generation
1- Create Video Task Submits an asynchronous text-to-video task using a Wan model and returns a task ID. Generation typically takes 1 to 5 minutes. Supports negative prompts, resolution selection, clip duration, prompt extension, and seed control. Result URLs expire after 24 hours.
Async Tasks
1- Get Task Status Retrieves the status and result of an asynchronous task created by Create Image Task or Create Video Task: PENDING, RUNNING, SUCCEEDED, FAILED, CANCELED, or UNKNOWN. On success, image tasks return image URLs and video tasks return the video URL.
Audio
2- Synthesize Speech Converts text to natural speech. qwen3-tts-flash supports 10+ languages, a large multilingual voice roster, and Chinese regional dialects with up to 600 characters of input. Generated audio is automatically saved to FlowRunner file storage with a permanent URL.
- Transcribe Audio Transcribes speech to text with qwen3-asr-flash from a public URL or base64 data URL in common formats. Detects the spoken language automatically, reports detected emotion, and supports biasing context for domain terms, an explicit language hint, and inverse text normalization for Chinese and English.
Models
1- List Models Lists all models available to your Model Studio API key, including chat, vision, embedding, and audio models with their IDs. Used to discover exact model IDs for the other actions.
Frequently Asked Questions
What can FlowRunner do with Qwen?
FlowRunner agents can run Chat Completion, Chat Completion (Advanced), and Analyze Image in Qwen, plus 8 more actions.
Does connecting Qwen to FlowRunner require OAuth?
No. Qwen connects to FlowRunner with an API key, no OAuth flow required.
Can Qwen trigger a FlowRunner workflow automatically?
Qwen doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with Qwen
$100 in credits. No card required. Connect in minutes.