> ## Documentation Index
> Fetch the complete documentation index at: https://phidatainc-studio-tools-doc.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# DeepSeek

> Use DeepSeek models with Agno agents.

DeepSeek is a platform for providing endpoints for Large Language models.
See their library of models [here](https://api-docs.deepseek.com/quick_start/pricing).

We recommend experimenting to find the best-suited model for your use-case. Here are some general recommendations:

* `deepseek-v4-flash` (default): fast V4 model, 1M context. Hybrid thinking + non-thinking.
* `deepseek-v4-pro`: flagship V4 model, 1M context. Hybrid thinking + non-thinking.

DeepSeek does not have rate limits. See their [docs](https://api-docs.deepseek.com/quick_start/rate_limit) for information about how to deal with slower responses during high traffic.

## Authentication

Set your `DEEPSEEK_API_KEY` environment variable. Get your key from [here](https://platform.deepseek.com/api_keys).

<CodeGroup>
  ```bash Mac theme={null}
  export DEEPSEEK_API_KEY=***
  ```

  ```bash Windows theme={null}
  setx DEEPSEEK_API_KEY ***
  ```
</CodeGroup>

## Example

Use `DeepSeek` with your `Agent`:

<CodeGroup>
  ```python agent.py theme={null}
  from agno.agent import Agent
  from agno.models.deepseek import DeepSeek

  agent = Agent(model=DeepSeek(id="deepseek-v4-flash"), markdown=True)

  # Print the response in the terminal
  agent.print_response("Share a 2 sentence horror story.")

  ```
</CodeGroup>

<Note> View more examples [here](/models/providers/native/deepseek/usage/basic). </Note>

## Parameters

| Parameter      | Type             | Default                      | Description                                                       |
| -------------- | ---------------- | ---------------------------- | ----------------------------------------------------------------- |
| `id`           | `str`            | `"deepseek-v4-flash"`        | The id of the DeepSeek model to use                               |
| `name`         | `str`            | `"DeepSeek"`                 | The name of the model                                             |
| `provider`     | `str`            | `"DeepSeek"`                 | The provider of the model                                         |
| `api_key`      | `Optional[str]`  | `None`                       | The API key for DeepSeek (defaults to `DEEPSEEK_API_KEY` env var) |
| `base_url`     | `str`            | `"https://api.deepseek.com"` | The base URL for the DeepSeek API                                 |
| `use_thinking` | `Optional[bool]` | `None`                       | Control thinking mode. See [Thinking mode](#thinking-mode).       |

`DeepSeek` extends the OpenAI-compatible interface and supports most parameters from the [OpenAI model](/models/providers/native/openai/completion/overview).

**Note**: DeepSeek supports JSON mode but not native `json_schema` structured outputs, so `supports_native_structured_outputs` is set to `False`. Use `use_json_mode=True` for structured output.

## Available Models

| Model id            | Notes                                                                |
| ------------------- | -------------------------------------------------------------------- |
| `deepseek-v4-flash` | Fast V4 model (default), 1M context. Hybrid thinking + non-thinking. |
| `deepseek-v4-pro`   | Flagship V4 model, 1M context. Hybrid thinking + non-thinking.       |

<Warning>
  The legacy ids `deepseek-chat` and `deepseek-reasoner` still work and route server-side to the V4 models (`deepseek-chat` to non-thinking, `deepseek-reasoner` to thinking), but you should migrate to `deepseek-v4-flash` / `deepseek-v4-pro`.
</Warning>

## Thinking mode

V4 models run with thinking enabled by default, so you get `reasoning_content` out of the box. Control it with the `use_thinking` flag:

| Value            | Behavior                                                     |
| ---------------- | ------------------------------------------------------------ |
| `None` (default) | V4 models think by default; legacy `deepseek-chat` does not. |
| `True`           | Force thinking on. The model returns `reasoning_content`.    |
| `False`          | Force thinking off for a faster, cheaper response.           |

For complex tasks, set `reasoning_effort="high"` or `reasoning_effort="max"`. While thinking is active, `temperature`, `top_p`, `presence_penalty` and `frequency_penalty` are ignored by the API.

```python theme={null}
from agno.agent import Agent
from agno.models.deepseek import DeepSeek

agent = Agent(
    model=DeepSeek(id="deepseek-v4-pro", reasoning_effort="max"),
    markdown=True,
)
```