​

Overview

• OpenAI-compatible by default - Chat Completions use /chat/completions, authenticated with a bearer token
• Optional Anthropic-compatible mode - Set use_anthropic_endpoints on a key (or override it per-alias) to route Chat Completions and the Responses API through /anthropic/v1/messages, authenticated with x-api-key, using the shared Anthropic request/response converters
• Streaming support - Server-Sent Events for chat, responses, and text completions, in both endpoint modes
• Tool calling - Function tools are supported on both the OpenAI-compatible and Anthropic-compatible paths
• Reasoning support - Reasoning parameters are mapped through the OpenAI converters by default, or the Anthropic converters when Anthropic-compatible mode is enabled
• Beta text completions - Text/FIM completions always use DeepSeek’s OpenAI-compatible /beta/completions endpoint, regardless of use_anthropic_endpoints

​

Supported Operations

​

Setup & Configuration

• Web UI
• config.json
• API
• Go SDK

1. Navigate to Models > Model Providers. Look for DeepSeek under Configured Providers. If it is missing, click on Add New Provider and select DeepSeek.
2. Click Add Key or edit an existing key.
3. Set a name for your key.
4. Paste your API key directly or use an environment variable (for example, env.DEEPSEEK_API_KEY).
5. Set Allowed Models to All Models (default) or the specific model allowlist you want this key to serve.
6. Leave Use Anthropic Endpoints off to use DeepSeek’s OpenAI-compatible endpoints (the default), or turn it on to route Chat Completions and the Responses API through DeepSeek’s Anthropic-compatible endpoint instead. See Anthropic-Compatible Endpoints below.
7. Save the provider configuration.

{
  "providers": {
    "deepseek": {
      "keys": [
        {
          "name": "deepseek-key-1",
          "value": "env.DEEPSEEK_API_KEY",
          "models": [
            "*"
          ],
          "weight": 1.0
        }
      ]
    }
  }
}

Refer to the API documentation for Provider Keys Management.

case schemas.DeepSeek:
    return []schemas.Key{{
        Name:   "deepseek-key-1",
        Value:  *schemas.NewSecretVar("env.DEEPSEEK_API_KEY"),
        Models: []string{"*"},
        Weight: 1.0,
    }}, nil

​

Anthropic-Compatible Endpoints (optional)

• Key-level - Sets the default endpoint mode for every request made with that key.
• Alias-level - Overrides the key-level default for a single alias, so one key can serve some aliases through the OpenAI-compatible endpoints and others through the Anthropic-compatible endpoint.

{
  "providers": {
    "deepseek": {
      "keys": [
        {
          "name": "deepseek-key-1",
          "value": "env.DEEPSEEK_API_KEY",
          "models": [
            "*"
          ],
          "weight": 1.0,
          "use_anthropic_endpoints": true
        }
      ]
    }
  }
}

{
  "model_id": "deepseek-v4-flash",
  "use_anthropic_endpoints": false
}

​

1. Chat Completions

​

Request Parameters

​

Authentication

​

Reasoning Parameter

• Default (OpenAI-compatible): Reasoning parameters follow the same conventions as the OpenAI provider (for example, reasoning.effort).
• use_anthropic_endpoints: true: Reasoning/thinking parameters are mapped through the Anthropic converters (reasoning → thinking), the same as the Anthropic provider. Reasoning effort is sent as output_config.effort (Anthropic’s own field placement), not nested under thinking.reasoning_effort as DeepSeek’s native API documents it.

​

Forced Tool Choice

• Default (OpenAI-compatible): Thinking is disabled when tool_choice is the generic "required" string (forcing some tool call, without pinning a specific one).
• use_anthropic_endpoints: true: Thinking is disabled when tool_choice pins a specific named function, for both Chat Completions and the Responses API. tool_choice: "required"/"any" is left untouched in this mode, since DeepSeek’s Anthropic-compatible endpoint accepts that combination with thinking on.

​

Extra Parameters

​

2. Responses API

• Default (OpenAI-compatible): Responses requests fall back to Chat Completions, the same conversion pattern used by other OpenAI-compatible-only providers:
  ResponsesRequest → ChatRequest → Response conversion
  
• use_anthropic_endpoints: true: Responses requests are sent natively to DeepSeek’s Anthropic-compatible endpoint at /anthropic/v1/messages — there is no internal conversion to Chat Completions. Both non-streaming and streaming Responses requests build an Anthropic-format request body directly from the BifrostResponsesRequest and convert the response back to Bifrost’s Responses format.

​

3. Text Completions

​

4. Text Completions Streaming

​

5. List Models

​

Unsupported Features

​

Caveats