Use ModelFold in Dify

中文

ModelFold is one OpenAI-style API for chat, image, video, and audio models. In Dify, add it as the OpenAI-API-compatible model provider and your chat apps can call ModelFold models with your own API key.

Site: https://ai.modelfold.online

1. Create an API key

1. Sign up at https://ai.modelfold.online.
2. Open Account → API Keys: https://ai.modelfold.online/account
3. Create a key. It looks like mf- followed by a random string. Copy it once and store it yourself. This repository never contains a real key.

New accounts receive signup credits (the site default is 100; the amount that lands in the wallet is what you can spend). Every successful call deducts credits. If the upstream model fails, ModelFold refunds that call.

2. What to point Dify at

API base URL	https://ai.modelfold.online/v1
Chat	POST /v1/chat/completions
List models	GET /v1/models
Images	POST /v1/images/generations
Auth	Authorization: Bearer <your key>

Dify appends /chat/completions itself. Put the base URL only. Do not paste /chat/completions into the form.

Chat on this base URL accepts model, messages, max_tokens (default 1024), and temperature. It always returns one JSON object (object: chat.completion). It does not emit server-sent events. It does not accept image parts inside messages, tools / function calls, or stream: true as a real stream. Leave those switches off.

export MODELFOLD_API_KEY="your-key"
curl -sS https://ai.modelfold.online/v1/models \
  -H "Authorization: Bearer $MODELFOLD_API_KEY"

curl -sS https://ai.modelfold.online/v1/chat/completions \
  -H "Authorization: Bearer $MODELFOLD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-6-luna","messages":[{"role":"user","content":"Reply with the single word pong."}],"max_tokens":32}'

Use a chat model id from GET /v1/models where modality is chat. An image id on this endpoint returns 400 unsupported_model.

3. Install the provider

Same steps on Dify Cloud (https://cloud.dify.ai) and on a self-hosted Dify. A self-hosted instance only needs outbound HTTPS to ai.modelfold.online.

1. In the workspace, open Settings → Model Provider. Some current builds label this Integrations → Model Provider. You need owner or admin on the workspace.
2. If OpenAI-API-compatible is not on the page, install it from the plugin marketplace. The publisher is langgenius, plugin id openai_api_compatible.
3. On that card, choose Add Model.

4. Fill in the model

Add one row per model id. Suggested first model: gpt-6-luna (lowest chat price in the catalog below).

Field	Value
Model type	LLM
Model name	gpt-6-luna (exact id)
Model display name	optional, for example GPT-6 Luna (ModelFold)
API Key	the key only, without the word Bearer
API Base URL	https://ai.modelfold.online/v1
model name for API endpoint	leave empty
Completion mode	Chat
Model context size	128000
Upper bound for max tokens	4096
Function Call Type	Not Support
Stream function calling	Not Support
Vision Support	Not Support
Video / Audio / Document Support	Not Support
Thinking Mode Support	Not supported
Web Search Support	Not supported
Structured Output	Not supported
Compatibility mode	OpenAI compatible (strict)
API Type	Chat Completions API (/chat/completions)
Token parameter name	Use max_tokens
Include Usage in Stream	Disabled
Stream mode auth	Not use
User Identity Support	Not Support

Notes on those fields:

• Model context size is only Dify's history budget. GET /v1/models does not return a context window. 128000 is a working cap so Dify trims long chats. A smaller number is fine.
• Upper bound for max tokens is the highest value the app's Max Tokens slider can reach. ModelFold pre-authorizes credits from max_tokens. On a new account, keep the app slider at 512 or 1024 so a large slider does not return 402 before the model runs.
• Token parameter name must be Use max_tokens. On Auto, model names that start with gpt-5, gpt-6, o1, or o3 make this plugin send max_completion_tokens. ModelFold reads max_tokens and ignores the other name (the call still runs, at the default of 1024).
• API Type must stay on Chat Completions. Responses API calls /v1/responses, which this base URL does not serve.
• Function calling and vision are off because this chat endpoint does not implement them. Do not use the model in an Agent app that expects tools, and do not attach images to the chat.

Save. The plugin's save check sends a short non-streaming ping (Stream mode auth = Not use). That check matches what ModelFold implements. A failed save is almost always the URL or the key; see Troubleshooting.

Repeat Add Model for each id you want in the picker. The key and base URL are the same; only the model name changes.

Chat model ids

Checked against GET /v1/models on 2026-10-27 (35 chat models). Prices are credits per 1,000,000 tokens from the ModelFold docs the same day. The catalog changes; list models again before you depend on a name. Model pages: https://ai.modelfold.online/models/<id>.

Model id	Input / 1M	Output / 1M
claude-fable-5	1200	6000
claude-haiku-4-5	82.5	427.5
claude-haiku-5-5	18	90
claude-opus-4-5	427.5	2145
claude-opus-4-6	427.5	2145
claude-opus-4-7	427.5	2145
claude-opus-4-8	600	3000
claude-opus-5	600	3000
claude-opus-5-5	480	2400
claude-sonnet-4-5	255	1282.5
claude-sonnet-4-6	255	1282.5
claude-sonnet-5	255	1282.5
claude-sonnet-5-5	240	1200
deepseek-v4-1-flash	36	142.5
gemini-3-pro	150	1050
gemini-3.1-pro	150	1050
gemini-3-6-flash	67.5	337.5
gemini-3-6-flash-openai	67.5	337.5
gemini-3-7-flash	67.5	337.5
gemini-3-7-flash-openai	67.5	337.5
gemini-3-8-flash	67.5	337.5
gemini-3-8-flash-openai	67.5	337.5
gemini-4-argon	240	1200
gemini-4-argon-openai	240	1200
gpt-5-2	131.256	1050
gpt-5-6-luna	16.8	100.8
gpt-5-6-sol	420	2520
gpt-5-6-terra	168	1008
gpt-6-1-sol	180	900
gpt-6-astra	840	4200
gpt-6-luna	9	45
gpt-6-sol	180	900
grok-4-6	240	720
grok-4-7	240	720
kimi-k3	720	3600

Image models are a different endpoint

The OpenAI-API-compatible plugin can add LLM, embeddings, rerank, speech-to-text, and text-to-speech. It has no image-generation type, and ModelFold does not serve embeddings, rerank, or speech on /v1. Do not add image ids as LLM models.

Image generation is POST https://ai.modelfold.online/v1/images/generations with model, prompt, and optional resolution or image_url. On 2026-10-27 the live list had 64 image models. Examples:

Model id	Starting credits
google/imagen4-fast	6
google/imagen4	12
gpt-image/1.5-text-to-image	6
flux-2/pro-text-to-image	8

Call that URL from a Dify HTTP Request node (POST, JSON body, header Authorization: Bearer <key>, read timeout 120 seconds). The response shape is { "ok": true, "data": [{ "url": "https://..." }] }.

5. Use the model

Chat assistant

1. Studio → Create from blank → Chatbot.
2. In the model picker on the prompt panel, choose OpenAI-API-compatible / gpt-6-luna (or the display name you set).
3. Set Max Tokens to 512.
4. Publish, open Preview, and send a short question.

Workflow LLM node

1. Studio → Create from blank → Workflow.
2. Put an LLM node between Start and End.
3. Select the same ModelFold model. User prompt: the start-node question variable. Leave vision off.
4. Run with a short question.

You can import examples/modelfold-llm.workflow.yml (Studio → Import DSL). The LLM node expects a provider model whose name is exactly gpt-6-luna. If the node shows the model as missing, open it and select the model you added. The file has no API key.

Workflow that does not depend on streaming

Dify usually asks the provider plugin to stream when a chat or an LLM node runs. ModelFold answers with one JSON body, so that path can fail after the model saves successfully. The HTTP workflow below is the call that matches the API: one POST, no stream.

1. Import examples/modelfold-hello.workflow.yml.
2. Open the app → environment variables → set MODELFOLD_API_KEY to your key.
3. Run. Input is question. Outputs are status_code and body. The assistant text is choices[0].message.content inside body.

To change models, edit the JSON body and replace gpt-6-luna with another chat id.

6. Troubleshooting

What you see	What to change
401 / unauthorized	The key is missing, revoked, or has extra text. Paste the key only. In the HTTP example the header must be Bearer plus a space plus the key.
404, URL contains /v1/v1	The plugin added /v1 again. Set API Base URL to https://ai.modelfold.online.
404, URL is /chat/completions with no /v1	An older plugin joined the path and dropped v1. Set API Base URL to https://ai.modelfold.online/v1/ (trailing slash).
404, URL ends with /chat/completions/chat/completions	You pasted the full path. Use only https://ai.modelfold.online/v1.
400 unsupported_model	The model name is not a chat id, or an image id was used on the chat endpoint. Copy the id from GET /v1/models.
402 insufficient credits	The wallet cannot cover the pre-authorization. Lower Max Tokens or add credits at https://ai.modelfold.online/account.
429	60 requests per minute for the account. Wait and retry.
Save works, chat or LLM node errors while parsing a stream	Expected with today's chat API: Dify sent stream: true, ModelFold returned one JSON object. Use examples/modelfold-hello.workflow.yml. Keep Include Usage in Stream = Disabled and Stream mode auth = Not use.
Agent tools never run	Function calling is not supported on this endpoint. Use a workflow LLM or HTTP node for plain text.
Empty or very long answers on gpt-5* / gpt-6*	Token parameter name is still Auto, so the plugin sent max_completion_tokens. Set it to Use max_tokens, then save the model again.

Other error types from the gateway: invalid_messages (400, empty messages), upstream_failed (502, refunded), upstream_unavailable (503).

7. Billing

Chat is settled on real token counts after a pre-authorization: credits = ceil((input tokens × input price + output tokens × output price) / 1,000,000). Image and video calls charge per generation. Prices and the credit packs are on https://ai.modelfold.online. A short gpt-6-luna reply often costs 1 credit.