Supported upstreams

An upstream is the provider API the DIAL OpenAI Adapter calls for a model. This reference shows, for each supported provider and model type, the DIAL Core model entry that connects it, and the adapter variable a model must be listed in, where one is needed. It is for platform engineers who add models to DIAL; it assumes you know the adapter overview and Models in DIAL Core.

In the samples, <ADAPTER> is the adapter's base URL, for example http://adapter-dial-openai:5000. The adapter receives the upstream URL and key from Core in the X-UPSTREAM-ENDPOINT and X-UPSTREAM-KEY headers.

Chat models

Azure OpenAI

Use the v1 API, which carries the deployment name in the request body instead of the URL:

"<DIAL_MODEL_ID>": {
  "type": "chat",
  "overrideName": "<AZURE_DEPLOYMENT>",
  "endpoint": "<ADAPTER>/openai/deployments/<AZURE_DEPLOYMENT>/chat/completions",
  "upstreams": [
    {
      "endpoint": "https://<AZURE_OPENAI_SERVICE>.openai.azure.com/openai/v1/chat/completions",
      "key": "<OPTIONAL_API_KEY>"
    }
  ]
}

Without a key, the adapter authenticates with Azure's DefaultAzureCredential, such as a managed identity.
The legacy dated API, .../openai/deployments/<AZURE_DEPLOYMENT>/chat/completions, still works. Keep the same deployment name in endpoint and in every upstream, or prefer the v1 API to avoid that mismatch.
To reach features only the Responses API offers, such as reasoning summaries and the web search tool, point the upstream at https://<AZURE_OPENAI_SERVICE>.openai.azure.com/openai/v1/responses. The adapter still serves the model through Chat Completions. See Capabilities.

OpenAI Platform

"<DIAL_MODEL_ID>": {
  "type": "chat",
  "overrideName": "<OPENAI_MODEL>",
  "endpoint": "<ADAPTER>/openai/deployments/<OPENAI_MODEL>/chat/completions",
  "upstreams": [
    { "endpoint": "https://api.openai.com/v1/chat/completions", "key": "<OPENAI_API_KEY>" }
  ]
}

The key is required. The legacy Completions API, https://api.openai.com/v1/completions, is also supported, for Azure-style and OpenAI Platform endpoints.

Azure AI Foundry

Models such as gpt-oss-120b or Mistral-Large-2411 deploy only to Azure AI Foundry. Use its Azure OpenAI endpoint, https://<FOUNDRY_SERVICE>.openai.azure.com/openai/v1/chat/completions, or the legacy model inference endpoint, https://<FOUNDRY_SERVICE>.services.ai.azure.com/models/chat/completions.

Amazon Bedrock (OpenAI models)

"<DIAL_MODEL_ID>": {
  "type": "chat",
  "overrideName": "openai.gpt-5.4",
  "endpoint": "<ADAPTER>/openai/deployments/openai.gpt-5.4/chat/completions",
  "upstreams": [
    {
      "endpoint": "https://bedrock-mantle.<AWS_REGION>.api.aws/openai/v1/chat/completions",
      "key": "<OPTIONAL_BEDROCK_BEARER_TOKEN>"
    }
  ]
}

Two endpoint formats are recognized: Bedrock Mantle (bedrock-mantle.<region>.api.aws) takes the plain model ID, such as openai.gpt-5.4; Bedrock Runtime (bedrock-runtime.<region>.amazonaws.com) expects it prefixed with the region, such as us.openai.gpt-5.4.
The region comes from the upstream URL. Each upstream can carry its own credentials, so one model can be balanced across AWS accounts.
The adapter uses the first credentials it finds: a bearer token in key; static keys aws_access_key_id and aws_secret_access_key (with optional aws_session_token); a role, aws_assume_role_arn; or the AWS SDK's default credential chain. Set the last three per upstream in extra_data, or for the whole adapter with the AWS_* variables in Configuration.

Mistral Platform

Upstream https://api.mistral.ai/v1/chat/completions, with the model name from the Mistral Platform. List the deployment in MISTRAL_DEPLOYMENTS. Magistral models return reasoning; see Reasoning content.

Alibaba Cloud Model Studio

Model Studio serves the Qwen series and third-party models such as DeepSeek, Kimi, and GLM through an OpenAI-compatible API:

"upstreams": [
  {
    "endpoint": "https://<WORKSPACE_ID>.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions",
    "key": "<API_KEY>",
    "extra_data": { "vendor": "alibaba-cloud" }
  }
]

The host differs by region. extra_data.vendor turns on prompt caching; without it, the adapter treats the upstream as a plain OpenAI one.

vLLM

Upstream <VLLM_ORIGIN>/v1/chat/completions. List the deployment in VLLM_DEPLOYMENTS, so the adapter uses vLLM's own token counting. For the Qwen3-ASR speech-to-text model served by vLLM, list it in QWEN3_ASR_VLLM_DEPLOYMENTS.

Claude and other Anthropic Messages providers

The adapter serves models behind the Anthropic Messages API: Claude in Azure AI Foundry, the Anthropic API, and compatible providers such as Fireworks or OpenRouter. The Claude-specific behavior comes from the bundled Anthropic adapter package.

"<DIAL_MODEL_ID>": {
  "type": "chat",
  "endpoint": "<ADAPTER>/openai/deployments/<CLAUDE_MODEL>/chat/completions",
  "upstreams": [
    { "endpoint": "https://<FOUNDRY_SERVICE>.services.ai.azure.com/anthropic/v1/messages", "key": "<OPTIONAL_API_KEY>" }
  ],
  "defaults": { "max_tokens": 4096 }
}

For the Anthropic API, use https://api.anthropic.com/v1/messages with a key. Without a key, Azure AI Foundry falls back to Microsoft Entra ID authentication. Claude needs max_tokens; set a default per model in defaults, as above, rather than through CLAUDE_DEFAULT_MAX_TOKENS.

To serve the same model to Anthropic-native clients such as Claude Code, see Anthropic Messages passthrough.

Image, video, and speech models

These models are configured as chat deployments in DIAL.

Model type	Upstream endpoint	Adapter variable
DALL·E 3	`https://<SERVICE>.openai.azure.com/openai/deployments/dall-e-3/images/generations`	`DALLE3_DEPLOYMENTS`
GPT Image 1	`.../openai/deployments/gpt-image-1/images/generations`	`GPT_IMAGE_1_DEPLOYMENTS`
Sora	`https://<SERVICE>.openai.azure.com/openai/v1/video/generations`	—
Sora 2, text-to-video and image-to-video	`https://<SERVICE>.openai.azure.com/openai/v1/videos`	—
Text-to-speech: `tts`, `tts-hd`, `gpt-4o-mini-tts`	`https://<SERVICE>.openai.azure.com/openai/v1/audio/speech`	—
Speech-to-text	`https://<SERVICE>.openai.azure.com/openai/v1/audio/transcriptions`	—
Text-to-speech on vLLM-Omni, such as MOSS-TTS	`<VLLM_ORIGIN>/v1/audio/speech`	`VLLM_DEPLOYMENTS`

The Azure Audio API version is set by AUDIO_AZURE_API_VERSION. Image and video models take their parameters, such as size or duration, in custom_fields.configuration; see Configurable models.

Responses API models

To serve a model through the OpenAI Responses API as well, add responsesEndpoint next to the required endpoint:

"<DIAL_MODEL_ID>": {
  "type": "chat",
  "overrideName": "<MODEL>",
  "endpoint": "<ADAPTER>/openai/deployments/<MODEL>/chat/completions",
  "responsesEndpoint": "<ADAPTER>/openai/v1/responses",
  "upstreams": [
    {
      "id": "1",
      "endpoint": "https://<AZURE_OPENAI_SERVICE>.openai.azure.com/openai/v1/chat/completions",
      "responsesEndpoint": "https://<AZURE_OPENAI_SERVICE>.openai.azure.com/openai/v1/responses"
    }
  ]
}

Upstream	Responses endpoint
Azure OpenAI	`https://<SERVICE>.openai.azure.com/openai/v1/responses`
OpenAI Platform	`https://api.openai.com/v1/responses`
Amazon Bedrock	`https://bedrock-mantle.<region>.api.aws/openai/v1/responses`; use `overrideName` with the Bedrock model ID
Alibaba Cloud Model Studio	`https://<WORKSPACE_ID>.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/responses`, with `extra_data.vendor`

Limitations: WebSocket mode is not supported, references to DIAL files are not supported, and passing context from a previous response works only for models with one upstream. DIAL Core needs each upstream to have an id; see Responses API in DIAL Core.

Embedding models

"<DIAL_MODEL_ID>": {
  "type": "embedding",
  "endpoint": "<ADAPTER>/openai/deployments/<EMBEDDING_DEPLOYMENT>/embeddings",
  "upstreams": [
    { "endpoint": "https://<AZURE_OPENAI_SERVICE>.openai.azure.com/openai/v1/embeddings", "key": "<OPTIONAL_API_KEY>" }
  ]
}

Upstream	Endpoint	Adapter variable
Azure OpenAI	`.../openai/v1/embeddings`, or the legacy `.../openai/deployments/<id>/embeddings`	—
OpenAI Platform	`https://api.openai.com/v1/embeddings`	—
Azure AI Vision, multimodal	`https://<VISION_SERVICE>.cognitiveservices.azure.com`	`AZURE_AI_VISION_DEPLOYMENTS`
vLLM	`.../v1/embeddings` or `.../pooling`, depending on the model	`VLLM_DEPLOYMENTS`

Multimodal models take text in input and images in custom_input, as {"type": "image/png", "url": "..."} objects:

curl -X POST "<DIAL_CORE>/openai/deployments/<DIAL_MODEL_ID>/embeddings" \
  -H "Api-Key: <DIAL_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{"input": ["cat", "fish"], "custom_input": [{"type": "image/png", "url": "https://example.com/cat.png"}]}'

Other OpenAI-compatible servers

A model that does not fall into any category above is treated as a text chat or text embeddings model on the OpenAI API, so any OpenAI-compatible server works as an upstream. The ai-dial-samples connect Ollama this way, with http://ollama:11434/v1/chat/completions. List Databricks chat models in DATABRICKS_DEPLOYMENTS, so the adapter relies on the provider's own token usage.

Chat models​

Azure OpenAI​

OpenAI Platform​

Azure AI Foundry​

Amazon Bedrock (OpenAI models)​

Mistral Platform​

Alibaba Cloud Model Studio​

vLLM​

Claude and other Anthropic Messages providers​

Image, video, and speech models​

Responses API models​

Embedding models​

Other OpenAI-compatible servers​

Further reading​