DeepInfra - GoModel

DeepInfra serves open and proxied models through an OpenAI-compatible API at https://api.deepinfra.com/v1/openai. GoModel routes these endpoints to it unchanged:

GoModel endpoint	DeepInfra models
`/v1/chat/completions`	chat models (`Qwen/Qwen3-32B`, `openai/gpt-oss-120b`, …)
`/v1/responses`	chat models, served by DeepInfra’s native Responses API
`/v1/embeddings`	embedding models (`Qwen/Qwen3-Embedding-0.6B`, `BAAI/bge-m3`, …)
`/v1/images/generations`, `/v1/images/edits`	image models (`black-forest-labs/FLUX-1-schnell`, `Bria/fibo_edit`, …)
`/v1/audio/speech`	TTS models (`hexgrad/Kokoro-82M`, …)
`/v1/audio/transcriptions`, `/v1/audio/translations`	STT models (`openai/whisper-large-v3-turbo`, …)
`/v1/files`, `/v1/batches`	DeepInfra’s batch API

Configure

Create an API key in the DeepInfra dashboard and set:

DEEPINFRA_API_KEY=...

Or in config.yaml:

providers:
  deepinfra:
    type: deepinfra
    api_key: "${DEEPINFRA_API_KEY}"
    # base_url defaults to "https://api.deepinfra.com/v1/openai".

Override the base URL or pin a model list only when you need to:

DEEPINFRA_BASE_URL=https://api.deepinfra.com/v1/openai
DEEPINFRA_MODELS=Qwen/Qwen3-32B,Qwen/Qwen3-Embedding-0.6B

Models

GoModel reads DeepInfra’s live /models listing, including each model’s kind (chat, embedding, image, speech, transcription), context length, reasoning and vision tags, and prices. Video generation models are left out because GoModel has no video endpoint for DeepInfra. Model IDs contain a slash; use them verbatim, or prefix the provider name to pin the route:

curl http://localhost:8080/v1/chat/completions \
  -H "Authorization: Bearer $GOMODEL_MASTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "deepinfra/Qwen/Qwen3-32B", "messages": [{"role": "user", "content": "Hello"}]}'

Both reasoning_effort and reasoning: {"effort": ...} are honored by DeepInfra’s reasoning models, and none turns thinking off on hybrid models such as Qwen3.

Behavior notes

Responses are stateless. DeepInfra keeps no responses, so store: true has no upstream effect. With a GoModel response store configured, previous_response_id chains still work: the gateway replays the stored history before dispatch. Without one, DeepInfra rejects previous_response_id and conversation with an invalid-request error.
Structured Responses output goes through chat. DeepInfra’s native Responses endpoint rejects every text setting (text.format, text.verbosity), so GoModel serves those requests through chat completions, where json_schema and json_object work.
Speech defaults to MP3. DeepInfra answers WAV when response_format is omitted; GoModel sends OpenAI’s default (mp3) instead. Set response_format explicitly to get another format.
Image pricing is per image unit. DeepInfra lists image prices per image unit, and GoModel records that rate as the per-image price. Cost tracking is approximate for models DeepInfra bills by size or step count.
Passthrough is opt-in. Add deepinfra to ENABLED_PASSTHROUGH_PROVIDERS to expose /p/deepinfra/..., which forwards any path under the base URL with the gateway’s DeepInfra key.

​Configure

​Models

​Behavior notes

Configure

Models

Behavior notes