Cohere - GoModel

GoModel translates OpenAI-compatible Chat Completions, Responses, Embeddings, and Audio Transcription requests to Cohere’s native v2 API. Model discovery uses Cohere’s model catalog and exposes chat, embedding, and transcription models through /v1/models.

Configure

COHERE_API_KEY=...
# COHERE_BASE_URL=https://api.cohere.com   # optional override

Or in config.yaml:

providers:
  cohere:
    type: cohere
    api_key: "${COHERE_API_KEY}"

Multiple keys use the standard COHERE_API_KEY_2, COHERE_API_KEY_3, … rotation convention. You can restrict or supplement model discovery with COHERE_MODELS.

Chat and Responses

Use a Cohere Command model with either OpenAI-compatible endpoint:

curl http://localhost:8080/v1/chat/completions \
  -H "Authorization: Bearer $GOMODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "command-a-plus-05-2026",
    "messages": [{"role": "user", "content": "Explain vector search briefly"}]
  }'

Chat streaming, system and developer messages, image URL parts, JSON response formats, tool definitions, tool calls, and tool results are translated to the Cohere v2 format. /v1/responses uses the same chat adapter. Cohere reasoning models enable thinking by default. To control Cohere’s native reasoning behavior, include its thinking object as an extension:

{
  "model": "command-a-reasoning-08-2025",
  "messages": [{"role": "user", "content": "Solve this carefully..."}],
  "thinking": {"type": "enabled", "token_budget": 4000}
}

Thinking output is returned as the OpenAI-compatible reasoning_content extension in both normal and streaming responses.

Audio transcription

Cohere audio uses a dedicated speech-to-text model and endpoint; Command chat models do not accept audio content mixed into Chat Completions or Responses. Transcribe audio through the OpenAI-compatible endpoint:

curl http://localhost:8080/v1/audio/transcriptions \
  -H "Authorization: Bearer $GOMODEL_API_KEY" \
  -F "model=cohere/cohere-transcribe-03-2026" \
  -F "language=en" \
  -F "response_format=json" \
  -F "file=@recording.wav"

The Cohere transcription API requires language and returns JSON. GoModel accepts the OpenAI-compatible response_format=json field but omits it from the native request because JSON is Cohere’s fixed response format. Cohere does not support text-to-speech, prompt hints, timestamp granularities, or streaming transcription.

Embeddings

GoModel supplies Cohere’s required embedding fields while retaining the OpenAI request shape:

curl http://localhost:8080/v1/embeddings \
  -H "Authorization: Bearer $GOMODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "embed-v4.0",
    "input": ["first document", "second document"],
    "dimensions": 1024,
    "input_type": "search_document"
  }'

input_type defaults to search_document. Set it to search_query, classification, or clustering when that better describes the input. encoding_format supports float (the default) and base64. Cohere-native truncate, max_tokens, and priority fields are also passed through.

Native passthrough

Add cohere to server.enabled_passthrough_providers, then use paths such as /p/cohere/v2/chat or /p/cohere/v2/rerank to call Cohere-native APIs that do not have an OpenAI-compatible GoModel surface. See Cohere’s Chat API, Embed API, Audio Transcription API, and reasoning guide for native fields and model-specific capabilities.

​Configure

​Chat and Responses

​Audio transcription

​Embeddings

​Native passthrough

Configure

Chat and Responses

Audio transcription

Embeddings

Native passthrough