Overview

DeepSeek is an OpenAI-compatible provider with a dedicated Bifrost provider implementation for DeepSeek’s endpoint layout. Bifrost uses the shared OpenAI-compatible request and response converters, while preserving DeepSeek-specific extra parameters and routing text completions to the beta FIM endpoint. Key characteristics:

OpenAI-compatible chat - Chat Completions use /chat/completions
Streaming support - Server-Sent Events for chat and text completions
Tool calling - Function tools are passed through using the OpenAI-compatible schema
Reasoning support - Reasoning parameters and DeepSeek extra parameters are forwarded
Responses API - Supported by converting Responses requests to Chat Completions internally
Beta text completions - Text/FIM completions use DeepSeek’s /beta/completions endpoint

Supported Operations

Operation	Non-Streaming	Streaming	Endpoint
Chat Completions	✅	✅	`/chat/completions`
Responses API	✅	✅	`/chat/completions`
Text Completions	✅	✅	`/beta/completions`
List Models	✅	-	`/models`
Embeddings	❌	❌	-
Image Generation	❌	❌	-
Speech (TTS)	❌	❌	-
Transcriptions (STT)	❌	❌	-
Files	❌	❌	-
Batch	❌	❌	-

Unsupported Operations (❌): Embeddings, Image Generation, Speech, Transcriptions, Files, Batch, cached content, containers, token counting, compaction, OCR, rerank, video, and passthrough are not supported by the upstream DeepSeek API through this provider. These return UnsupportedOperationError.

Setup & Configuration

Configure DeepSeek as a provider.

Web UI
config.json
API
Go SDK

Navigate to Models > Model Providers. Look for DeepSeek under Configured Providers. If it is missing, click on Add New Provider and select DeepSeek.
Click Add Key or edit an existing key.
Set a name for your key.
Paste your API key directly or use an environment variable (for example, env.DEEPSEEK_API_KEY).
Set Allowed Models to All Models (default) or the specific model allowlist you want this key to serve.
Save the provider configuration.

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

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

1. Chat Completions

Request Parameters

DeepSeek supports OpenAI-compatible chat completion parameters. For the full parameter reference and message conversion behavior, see OpenAI Chat Completions.

Filtered Parameters

Removed for DeepSeek compatibility:

prediction - OpenAI-specific predicted output
prompt_cache_key - OpenAI-specific prompt cache key
prompt_cache_retention - OpenAI-specific prompt cache retention
verbosity - Anthropic-specific
store - OpenAI-specific response storage
web_search_options - OpenAI-specific web search options

Reasoning Parameter

DeepSeek delegates through ToOpenAIChatRequest with provider-specific compatibility handling. Reasoning effort is normalized using the OpenAI-compatible provider convention, and DeepSeek V4 models preserve reasoning.effort: "max" when requested. Assistant-message reasoning details are stripped before sending follow-up messages because DeepSeek rejects reasoning_details in assistant messages.

Extra Parameters

DeepSeek enables passthrough extra parameters for chat and text completion requests. Provider-specific options such as DeepSeek thinking controls can be sent through extra_params without being dropped by Bifrost. DeepSeek supports standard OpenAI message types, tools, responses, and streaming formats. For details on message handling, tool conversion, responses, and streaming, refer to OpenAI Chat Completions.

2. Responses API

Bifrost converts Responses API format to Chat Completions internally, then converts the response back:

BifrostResponsesRequest
  → ToChatRequest()
  → ChatCompletion
  → ToBifrostResponsesResponse()

Same parameter support as Chat Completions with response format differences (output items instead of message content). Streaming Responses requests are also routed through Chat Completions streaming.

3. Text Completions

DeepSeek supports beta text/FIM completions through /beta/completions:

Parameter	Mapping
`prompt`	Sent as-is
`max_tokens`	max_tokens
`temperature`	temperature
`top_p`	top_p
`stop`	stop sequences
`extra_params`	Passed through to DeepSeek

Response returns choices[].text with completion text.

4. Text Completions Streaming

Streaming text completions use DeepSeek’s OpenAI-compatible SSE format on /beta/completions.

5. List Models

Lists available models from DeepSeek through /models.

Unsupported Features

Feature	Reason
Embedding	Not offered by DeepSeek API through this provider
Image Generation	Not offered by DeepSeek API through this provider
Speech/TTS	Not offered by DeepSeek API through this provider
Transcription/STT	Not offered by DeepSeek API through this provider
Batch Operations	Not offered by DeepSeek API through this provider
File Management	Not offered by DeepSeek API through this provider
Cached Content	Only Gemini and Vertex AI support cached content in Bifrost
Container Management	Not offered by DeepSeek API through this provider
Token Counting	Not offered by DeepSeek API through this provider
Rerank/OCR/Video	Not offered by DeepSeek API through this provider

Caveats

Default Base URL

Severity: Low Behavior: DeepSeek defaults to https://api.deepseek.com Impact: Custom DeepSeek-compatible deployments must override network_config.base_url Code: NewDeepSeekProvider sets the default base URL when no provider-level base URL is configured

Beta Text Completion Endpoint

Severity: Medium Behavior: Text completions are routed to /beta/completions Impact: FIM/text completion behavior follows DeepSeek’s beta API contract and may differ from standard OpenAI /completions Code: TextCompletion and TextCompletionStream use /beta/completions

User Field Size Limit

Severity: Low Behavior: User field > 64 characters is silently dropped Impact: Longer user identifiers are lost Code: SanitizeUserField enforces 64-char max in the shared OpenAI converter

Reasoning Details in Assistant Messages

Severity: Medium Behavior: Assistant-message reasoning details are removed before forwarding follow-up chat messages Impact: Prevents DeepSeek request failures when previous assistant turns contain reasoning metadata Code: stripReasoningDetails applies to DeepSeek in ToOpenAIChatRequest

​Overview

​Supported Operations

​Setup & Configuration

​1. Chat Completions

​Request Parameters

​Filtered Parameters

​Reasoning Parameter

​Extra Parameters

​2. Responses API

​3. Text Completions

​4. Text Completions Streaming

​5. List Models

​Unsupported Features

​Caveats

Overview

Supported Operations

Setup & Configuration

1. Chat Completions

Request Parameters

Filtered Parameters

Reasoning Parameter

Extra Parameters

2. Responses API

3. Text Completions

4. Text Completions Streaming

5. List Models

Unsupported Features

Caveats