API protocol
Model-specific scenarios and support boundaries
Preview guide
Paste into an AI assistant to help with integration. Documentation and examples only; no real credentials.
Direct answerConfirm the model and endpoint in the marketplace, then migrate paths, fields and task results; legacy release tags do not guarantee gateway availability.
Updated · Reviewed
Video family guides
See the family comparison and complete examples for model-specific parameters, media roles, code, polling and downloads. Wan, Kling, Seedance, HappyHorse, Gemini Omni and MiniMax have separate contracts. Conditional and unpublished scenarios are explicitly labeled.
Select the model and protocol first
A shared /v1/chat/completions or /v1/videos path can serve models with different values, duration, resolution and pricing rules. Choose the exact marketplace model, key group and effective endpoint, then copy that model's integration guide. This page records code contracts reviewed on November 8, 2026; it does not invent production group configuration.
Scenarios not published
Sora Remix and Qwen Omni have not completed live acceptance in this round. Their integration references and examples have been withdrawn. Use the published model and protocol guides.
Qwen Image 3.0: generation and editing
These rules are enforced for qwen-image-3.0 and qwen-image-3.0-* through the Ali adapter. Other compatible channels require their own checks. Use image generation for JSON and multipart edits for local file upload.
| Field | Gateway Ali conversion rule |
|---|---|
| image | Omit for text-to-image; JSON accepts a URL or URL array, with at most 3 references for editing |
| n | 1–6 outputs; not a universal image-model range |
| prompt_extend | Boolean; explicit false survives |
| prompt_extend_mode | direct or agent; agent is only allowed without reference images |
| seed | 0–2147483647; explicit zero survives |
| size | widthxheight converts to width*height; area 512²–2048² and aspect 1:8–8:1 |
| watermark, negative_prompt | Boolean and negative prompt, subject to channel support |
{"model":"qwen-image-3.0","prompt":"Keep the subject and replace the background with a beach","image":["https://example.com/input.png"],"n":1,"size":"1024x1024","prompt_extend":false,"prompt_extend_mode":"direct","seed":0,"watermark":false}
This is a JSON editing scenario for /v1/images/generations when the model route supports references; a local file path is not an upload. Read data[].url or decode b64_json. An empty result is not success. The Ali multipart path additionally limits each image to 10 MiB; other paths have their own body limits.
Veo, Seedance and MiniMax differences
| Model family | Routing and limits |
|---|---|
| Veo | Reference modes, image count, duration, orientation and 4K depend on the exact model and channel profile. Historical components/pro/4k suffixes are not gateway IDs to invent; use effective endpoint values |
| Seedance | Text, first/last frames, multimodal references and real-person assets are different workflows. Uploaded real-person assets must pass compliance to active before generation; see the asset guide |
| MiniMax-H3 | Unified videos, Remix and native /v2/video_generation or /v2/video_regeneration have distinct inputs. Follow the relevant video guide for resolution, source and context |
| FLUX | Native /flux/v1/{model} accepts only flux-3-image. Async delivery requires an accessible result URL; Base64 alone does not prove task completion |
Advanced text fields and state boundaries
The Responses reference includes instructions, previous_response_id, conversation, reasoning, store and context_management. Parsing a field only establishes a gateway representation, not universal model support. Do not combine conversation and previous_response_id. Upstream session continuation across channels is not guaranteed. Channel settings can filter service_tier and safety_identifier or disable store. Background execution and retrieval/cancellation are unimplemented and are not listed as supported features.
Enable optional fields incrementally after the minimum request works. Bound polling and check terminal success or failure. Local contract checks are not paid upstream acceptance tests.
Use cases
- Migrate legacy clients
- Verify model fields and support boundaries
API protocols
/v1/chat/completions/v1/videos/suno/v2/music
FAQ
Does a legacy released label establish support here?
No. Check gateway routing, field conversion, channel and model-group support.
Can I migrate by changing only the URL?
No. Check authentication, fields, submission responses, task IDs, terminal states and results.
Have all models passed live acceptance tests?
No. This is a code/documentation contract review, not paid upstream acceptance for every model.
兔子API