Skip to main content

Provider Capability Matrix

The gateway exposes a single OpenAI-compatible surface, but not every provider implements every capability. This page is a provider × capability matrix generated from the gateway's built-in provider registry (providers/providers_list.go), so it always reflects exactly what each provider actually advertises.

What the capabilities mean​

  • Chat — chat completions via POST /v1/chat/completions. Every provider supports this.
  • Stream — incremental token streaming with "stream": true (server-sent events).
  • Embed — vector embeddings for retrieval, search, and clustering.
  • Image — image generation from a text prompt.
  • Discovery — live model discovery, so GET /v1/models reflects the provider's current catalog rather than only the static built-in catalog.
  • Proxy — transparent passthrough of unhandled /v1/* endpoints straight to the provider, preserving provider-native request and response shapes.

Each row corresponds to a provider's virtual key. A ✅ means the provider declares that capability in the registry; a — means it does not.

The matrix​

ProviderVirtual KeyChatStreamEmbedImageDiscoveryProxy
AI21ai21✅✅———✅
Anthropicanthropic✅✅———✅
Azure Foundryazure-foundry✅✅———✅
Azure OpenAIazure-openai✅✅———✅
AWS Bedrockbedrock✅✅✅——✅
Cerebrascerebras✅✅——✅✅
Cloudflare Workers AIcloudflare✅✅✅——✅
Coherecohere✅✅✅——✅
Databricksdatabricks✅✅✅——✅
DeepInfradeepinfra✅✅———✅
DeepSeekdeepseek✅✅———✅
Fireworks AIfireworks✅✅✅—✅✅
Google Geminigemini✅✅✅——✅
Groqgroq✅✅———✅
Hugging Facehugging-face✅✅✅✅✅✅
Mistralmistral✅✅✅——✅
Moonshot AImoonshot✅✅———✅
Novita AInovita✅✅✅—✅✅
NVIDIA NIMnvidia-nim✅✅———✅
Ollamaollama✅✅———✅
Ollama Cloudollama-cloud✅✅——✅—
OpenAIopenai✅✅✅✅✅✅
OpenRouteropenrouter✅✅——✅✅
Perplexityperplexity✅✅——✅✅
Qwen (Alibaba)qwen✅✅———✅
Replicatereplicate✅——✅—✅
SambaNovasambanova✅✅———✅
Together AItogether✅✅✅——✅
Vertex AIvertex-ai✅✅✅——✅
xAI (Grok)xai✅✅——✅✅

Capability totals​

Across the 30 built-in providers:

  • Chat — 30 providers
  • Stream — 29 providers (every provider except Replicate)
  • Embed — 11 providers
  • Image — 3 providers (Hugging Face, OpenAI, Replicate)
  • Discovery — 9 providers
  • Proxy — 29 providers (every provider except Ollama Cloud)

Notes​

  • Chat is universal. All 30 providers implement chat completions, so the gateway can route any request to any of them under your chosen strategy.
  • Streaming is near-universal. Only Replicate does not stream, because its text and image models run as polled predictions rather than token streams.
  • Discovery vs. the static catalog. Providers without Discovery still return models — they fall back to the built-in model catalog used for cost estimation and GET /v1/models — but they will not reflect newly released provider models until the catalog is updated.
  • Proxy reach. Almost every provider allows passthrough of provider-native endpoints; Ollama Cloud is the exception and is reached through the gateway's normalized chat and discovery paths only.

Next steps​