Skip to main content

Choose a Provider Upstream

A provider upstream is the model service an AISIX gateway calls after resolving a caller-facing model alias. AISIX stores the upstream credential in a provider key and uses an adapter to send requests in the format that the upstream expects.

Choose the setup path for the endpoint and account you use, not only for the model family. Configure Gemini through Google AI Studio when you have an AI Studio API key. Use Google Vertex AI when the model is hosted in a Google Cloud project.

Every setup below supports both AISIX Cloud and the open-source AISIX gateway. Each provider guide contains complete configuration sections for both products followed by one shared verification procedure.

Setup Paths

Choose among the direct provider and managed-platform setups:

SetupUse WhenAdapter
OpenAIYou call models through an OpenAI API account.openai
Azure OpenAIAn Azure OpenAI deployment hosts the model.azure-openai
AnthropicYou call Claude through the Anthropic API.anthropic
AWS BedrockAWS Bedrock hosts the model or inference profile.bedrock
Gemini (Google AI Studio)You call Gemini with a Google AI Studio API key.openai
Google Vertex AIA Google Cloud project hosts Gemini or another Vertex publisher model.vertex
Qwen (Alibaba Cloud)You call Qwen through Alibaba Cloud Model Studio.openai
DeepSeekYou call models through the DeepSeek API.openai
GroqYou call Groq-hosted models.openai
Mistral AIYou call models through the Mistral API.openai
Together AIYou call models from the Together AI catalog.openai
OpenRouterYou call models from many vendors through one OpenRouter account.openai
Fireworks AIYou call Fireworks-hosted models.openai
PerplexityYou call the search-grounded Sonar models.openai
CohereYou call Cohere Command models, embeddings, or rerank.openai
CerebrasYou call open-weight models on Cerebras inference hardware.openai
Moonshot AI (Kimi)You call Kimi models through the Moonshot AI platform.openai
Zhipu AI (GLM)You call GLM chat models or CogVideoX video generation.openai
Hugging FaceYou call open-weight models through the Hugging Face Inference Providers router.openai
BasetenYou call Baseten Model APIs or a dedicated Baseten deployment.openai

Three more setups target providers whose video or rerank wire the gateway implements natively. The openai adapter shapes their chat-style and embeddings traffic, while the video-generation and rerank routes dispatch on the provider value itself. These providers are not listed on the models.dev catalog, so the dashboard suggests no model IDs for them, and cost metadata for budgets and usage reports is operator-supplied on the model alias — see each page's cost metadata section:

SetupUse WhenAdapter
JinaYou call Jina embedding models, or reranker models through the rerank route.openai
RunwayMLYou run Runway Gen or Runway-hosted Veo video tasks through the video route.openai
Volcengine Ark (Doubao)You call Doubao chat models or Seedance video generation through Volcengine Ark.openai

AISIX Cloud can also create provider keys for community catalog providers from models.dev. These providers use the openai adapter by default, but AISIX does not curate provider-specific request rewrites, response rewrites, or base URL quirks for them. The open-source gateway can use the same provider identifiers as labels, but you must set the adapter and base URL explicitly in resources.yaml. Use these pages when the provider is available in the catalog and you want a provider-specific setup guide:

SetupUse WhenAdapter
Amazon Nova APIYou call Nova models through the direct API-key-based Nova endpoint rather than Bedrock.openai
Cloudflare Workers AIYou call account-scoped Cloudflare Workers AI model IDs such as @cf/openai/gpt-oss-120b.openai
DatabricksYou call Databricks Model Serving endpoints in your workspace.openai
DeepInfraYou call DeepInfra-hosted open-weight models through its OpenAI-compatible API.openai
DigitalOcean Gradient AIYou call foundation models through DigitalOcean Serverless Inference.openai
Meta Llama APIYou call Llama models through Meta's direct OpenAI-compatible API.openai
MiniMaxYou call MiniMax models through the OpenAI-compatible API root.openai
ModelScopeYou call ModelScope API-Inference model IDs.openai
Nebius Token FactoryYou call open models hosted by Nebius Token Factory.openai
NVIDIA NIMYou call the NVIDIA hosted NIM API or NVIDIA-published model IDs.openai
Novita AIYou call Novita-hosted open models through its LLM API.openai
OVHcloud AI EndpointsYou call models through the OpenAI-compatible endpoint from OVHcloud.openai
SiliconFlowYou call SiliconFlow-hosted org-namespaced model IDs.openai
Snowflake CortexYou call models available in a Snowflake account through the Cortex REST API.openai
Weights & Biases InferenceYou call open models through W&B Inference.openai
xAIYou call Grok models through the OpenAI-compatible API from xAI.openai

For another public or private endpoint that accepts OpenAI-compatible chat-completions requests, choose the corresponding general setup:

SetupUse WhenAdapter
Other OpenAI-Compatible ProvidersA public provider exposes an OpenAI-compatible API and does not have a dedicated setup guide.openai
OllamaYou run models on a local or private Ollama server.openai
vLLMYou serve a private model with the OpenAI-compatible server provided by vLLM.openai
Bring Your Own EndpointYou operate another private OpenAI-compatible endpoint, such as SGLang or an internal proxy.openai

Each setup creates the same gateway resources: a provider key, a model alias, and a caller API key. The setup page provides the credential, upstream model identifier, base URL, and adapter details for that provider. Compare route support in Provider Compatibility, or review the upstream request formats in Adapter Protocol Families.