DeepSeek AI provider – OpenAI-compatible Chat Completions API.

Implementation

Uses built-in OpenAI-style encoding/decoding defaults. DeepSeek is fully OpenAI-compatible, so no custom request/response handling is needed.

Authentication

Requires a DeepSeek API key from https://platform.deepseek.com/

Configuration

# Add to .env file (automatically loaded)
DEEPSEEK_API_KEY=your-api-key

Examples

# Basic usage
ReqLLM.generate_text("deepseek:deepseek-chat", "Hello!")

# With custom parameters
ReqLLM.generate_text("deepseek:deepseek-reasoner", "Write a function",
  temperature: 0.2,
  max_tokens: 2000
)

# Streaming
ReqLLM.stream_text("deepseek:deepseek-chat", "Tell me a story")
|> Enum.each(&IO.write/1)

# With thinking mode enabled (default for reasoning models)
ReqLLM.generate_text("deepseek:deepseek-v4-pro", "Solve this complex problem",
  reasoning_effort: :high,
  provider_options: [
    thinking: %{type: "enabled"}
  ]
)

# With maximum reasoning effort for complex tasks
ReqLLM.generate_text("deepseek:deepseek-v4-pro", "Complex reasoning task",
  reasoning_effort: :xhigh
)

# Disable thinking mode
ReqLLM.generate_text("deepseek:deepseek-v4-pro", "Quick question",
  provider_options: [
    thinking: %{type: "disabled"}
  ]
)

Models

DeepSeek offers several models including:

• deepseek-chat - General purpose conversational model
• deepseek-reasoner - Reasoning and problem-solving
• deepseek-v4-flash - Fast reasoning model with lower latency
• deepseek-v4-pro - Latest reasoning model with thinking support

Thinking Mode

DeepSeek models support thinking mode for improved reasoning.

See DeepSeek Thinking Mode Guide for details.

Options

• thinking: %{type: "enabled"} - Enable thinking mode (default for reasoning models)
• thinking: %{type: "disabled"} - Disable thinking mode

Reasoning Effort

The reasoning_effort option controls the depth of reasoning. For compatibility, :low and :medium are mapped to "high", and :xhigh is mapped to "max".

Only "high" and "max" are the meaningful values sent to the API:

• :low → mapped to "high"
• :medium → mapped to "high"
• :high → "high" (default for thinking mode)
• :xhigh → mapped to "max" (maximum effort for complex tasks)

See https://platform.deepseek.com/docs for full model documentation.