ContextQMD
Back to Promptfoo

DeepSeek

site/docs/providers/deepseek.md

0.121.93.9 KB
Original Source

DeepSeek

DeepSeek provides an OpenAI-compatible API for their language models, with specialized models for both general chat and advanced reasoning tasks. The DeepSeek provider is compatible with all the options provided by the OpenAI provider.

Setup

  1. Get an API key from the DeepSeek Platform
  2. Set DEEPSEEK_API_KEY environment variable or specify apiKey in your config

Configuration

Basic configuration example:

yaml
providers:
  - id: deepseek:deepseek-chat
    config:
      temperature: 0.7
      max_tokens: 4000
      apiKey: YOUR_DEEPSEEK_API_KEY

  - id: deepseek:deepseek-reasoner # DeepSeek-R1 model
    config:
      max_tokens: 8000

Configuration Options

  • temperature
  • max_tokens
  • cost, inputCost, outputCost - Override promptfoo's pricing estimates (inputCost and outputCost take precedence over cost)
  • top_p, presence_penalty, frequency_penalty
  • stream
  • showThinking - Control whether reasoning content is included in the output (default: true, applies to deepseek-reasoner model)

Available Models

:::note

The API model names are aliases that automatically point to the latest versions: both deepseek-chat and deepseek-reasoner currently point to DeepSeek-V3.2, with the chat model using non-thinking mode and the reasoner model using thinking mode.

:::

deepseek-chat

  • General purpose model for conversations and content generation
  • 128K context window, up to 8K output tokens
  • Input: $0.028/1M (cache hit), $0.28/1M (cache miss)
  • Output: $0.42/1M

deepseek-reasoner

  • Specialized reasoning model with extended thinking capabilities
  • 128K context window, up to 64K output tokens (32K reasoning + final answer)
  • Input: $0.028/1M (cache hit), $0.28/1M (cache miss)
  • Output: $0.42/1M
  • Supports showing or hiding reasoning content through the showThinking parameter

:::warning

The reasoning model does not support temperature, top_p, presence_penalty, frequency_penalty, logprobs, or top_logprobs parameters. Setting these parameters will not trigger an error but will have no effect.

:::

Example Usage

Here's an example comparing DeepSeek with OpenAI on reasoning tasks:

yaml
providers:
  - id: deepseek:deepseek-reasoner
    config:
      max_tokens: 8000
      showThinking: true # Include reasoning content in output (default)
  - id: openai:o-1
    config:
      temperature: 0.0

prompts:
  - 'Solve this step by step: {{math_problem}}'

tests:
  - vars:
      math_problem: 'What is the derivative of x^3 + 2x with respect to x?'

Controlling Reasoning Output

The DeepSeek-R1 model (deepseek-reasoner) includes detailed reasoning steps in its output. You can control whether this reasoning content is shown using the showThinking parameter:

yaml
providers:
  - id: deepseek:deepseek-reasoner
    config:
      showThinking: false # Hide reasoning content from output

When showThinking is set to true (default), the output includes both reasoning and the final answer in a standardized format:

Thinking: <reasoning content>

<final answer>

When set to false, only the final answer is included in the output. This is useful when you want better reasoning quality but don't want to expose the reasoning process to end users or in your assertions.

See our complete example that benchmarks it against OpenAI's o1 model on the MMLU reasoning tasks.

API Details

  • Base URL: https://api.deepseek.com/v1
  • OpenAI-compatible API format
  • Full API documentation

See Also