DeepSeek V4 - Anthropic-Compatible API

##All APIs require Bearer Token authentication## **Get API Key**: Visit the [API Key Management Page](https://starmagic.ai/app/api-keys) to obtain your API Key **Add to request header**: ``` Authorization: Bearer YOUR_API_KEY ``` **Note**: Although Anthropic's native API uses the `x-api-key` header, EvoLink's `/v1/messages` uniformly uses Bearer Token authentication.

Authorization: Bearer YOUR_API_KEY

Model to invoke - `deepseek-v4-flash`: Fast general-purpose - `deepseek-v4-pro`: Deep reasoning **Tip**: Both models **have thinking enabled by default**, and responses always contain a `type="thinking"` content block; set `thinking.type="disabled"` explicitly to turn it off. An unspecified or unsupported model is automatically mapped to `deepseek-v4-flash`.

"deepseek-v4-flash"

Maximum number of tokens to generate (**required**) **Notes**: - The V4 series can reach up to **384,000** - Tokens produced by thinking also count toward the max_tokens limit

1024

List of conversation messages, alternating between user / assistant turns **Notes**: - At least one message must be included - The final message is typically `role=user` - `image` / `document` content types are not yet supported

[
  {
    "role": "user",
    "content": null
  }
]

System prompt, used to define the AI's role and behavior **Notes**: - Accepts a string or an array of strings - Unlike the `system` message on the OpenAI endpoint, the Anthropic endpoint uses a top-level `system` field

Sampling temperature **Notes**: - Range `[0.0, 2.0]` - Default 1; higher values are more diverse, lower values are more deterministic

1

Nucleus sampling threshold **Notes**: - Range `[0, 1]` - Do not adjust temperature and top_p simultaneously

1

Custom stop sequences **Notes**: - Generation stops when the model encounters any of these strings - Up to 4 entries (following Anthropic's specification)

[
  "string"
]

Whether to return via SSE streaming - `true`: Server-Sent Events streaming return - `false`: Return the complete response all at once (default)

false

Thinking mode control (V4) **Notes**: - **Enabled by default on both models** (`type=enabled`) - When enabled, the response `content` array includes a `type="thinking"` reasoning block (billed as output tokens) - **Note**: The API **ignores** Anthropic's native `budget_tokens` field; use `output_config.effort` to control depth - In multi-turn conversations, simply place the thinking block from the previous response back into the assistant `content` array verbatim (the Anthropic protocol is more lenient — missing thinking will not cause an error, but preserving the signature helps context consistency)

{
  "type": "enabled",
  "budget_tokens": 0
}

Output configuration (V4 extension) **Notes**: DeepSeek only supports the `effort` field

{
  "effort": "medium"
}

List of tool definitions **Notes**: - Follows the Anthropic tool definition specification - `input_schema` uses a JSON Schema object

[
  {
    "name": "string",
    "description": "string",
    "input_schema": {}
  }
]

Controls tool-calling behavior **Available types**: - `auto`: Model decides automatically (default when tools are provided) - `any`: Must call some tool (without specifying which) - `tool`: Must call the specified `name` - `none`: Forbid tool calls

{
  "type": "auto",
  "name": "string",
  "disable_parallel_tool_use": true
}