DeepSeek V4 - OpenAI-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 ```

Authorization: Bearer YOUR_API_KEY

Chat model name - `deepseek-v4-flash`: Fast general-purpose model, 1M context - `deepseek-v4-pro`: Deep reasoning model, excels at math, programming, and complex logic **Tip**: Both models **have `thinking` enabled by default**, and responses include `reasoning_content`. Set `thinking.type="disabled"` to turn it off and reduce output token cost. Both models share identical parameters.

"deepseek-v4-flash"

List of conversation messages, supports multi-turn dialogue Messages with different roles have different field structures; select the corresponding role to view details

[
  null
]

Thinking mode control (new in V4) **Notes**: - Controls the deep thinking (Chain of Thought) feature - **Enabled by default on both models** (`type=enabled`) - When enabled, the reasoning process is returned through `choices[].message.reasoning_content` and billed as output tokens ⚠️ **Multi-turn / tool-calling caveat**: If the current response includes `reasoning_content`, **the corresponding assistant message in the `messages` history of the next request must echo that field verbatim**, otherwise the API returns 400 `The reasoning_content in the thinking mode must be passed back to the API`. If you would rather not handle it, set `thinking.type="disabled"` explicitly for the whole session.

{
  "type": "enabled",
  "reasoning_effort": "medium"
}

Sampling temperature, controls randomness of output **Notes**: - Lower values (e.g., 0.2): More deterministic, more focused output - Higher values (e.g., 1.5): More random, more creative output - Default: 1

1

Nucleus sampling parameter **Notes**: - Controls sampling from tokens with cumulative probability - For example, 0.9 means sampling from tokens whose cumulative probability reaches 90% - Default: 1.0 (considers all tokens) **Suggestion**: Do not adjust temperature and top_p simultaneously

1

Limits the maximum number of tokens generated **Notes**: - The V4 series can reach up to **384,000 tokens** - When thinking is enabled, reasoning_tokens also count toward the max_tokens limit - If not set, the model decides the generation length on its own

4096

Frequency penalty, used to reduce repetitive content **Notes**: - Positive values penalize tokens based on their frequency in the already-generated text - The higher the value, the less likely repetition becomes - Default: 0 (no penalty)

0

Presence penalty, used to encourage new topics **Notes**: - Positive values penalize tokens based on whether they have already appeared in the text - The higher the value, the more the model tends to discuss new topics - Default: 0 (no penalty)

0

Specifies the response format **Notes**: - Set to `{"type": "json_object"}` to enable JSON mode - In JSON mode the model outputs valid JSON content - For best results, explicitly ask for JSON output in your system or user message

{
  "type": "text"
}

Stop sequences; generation stops when the model encounters any of these strings **Notes**: - Can be a single string or an array of strings - Up to 16 stop sequences are supported

Whether to stream the response - `true`: Stream response; returns content chunk by chunk in real time via SSE (Server-Sent Events) - `false`: Wait for the full response and return it at once (default)

false

Streaming response options Only effective when `stream=true`

{
  "include_usage": true
}

List of tool definitions for Function Calling **Notes**: - Up to 128 tool definitions are supported - Each tool must define a name, description, and parameter schema

[
  {
    "type": "function",
    "function": {
      "name": "string",
      "description": "string",
      "parameters": {},
      "strict": false
    }
  }
]

Controls tool-calling behavior **Options**: - `none`: Do not call any tool - `auto`: Let the model decide whether to call a tool (default when tools are provided) - `required`: Force the model to call one or more tools - Object form `{"type":"function","function":{"name":"xxx"}}`: Call the specified tool **Default**: `none` when no tools are provided, `auto` when tools are provided

Whether to return token log probabilities **Notes**: - When set to `true`, the response includes log probability information for each token

false

Return log probabilities of the top N tokens **Notes**: - Requires `logprobs` to be `true` - Range: `[0, 20]`

0

Token bias map **Notes**: - Keys are token IDs in the tokenizer; values are bias values between -100 and 100 - -100 completely bans the token, 100 forces it to be generated - Typical values in the range -1 to 1 already produce observable effects

{}

Number of chat completion choices to generate for each input message **Notes**: - Default 1; if set to N, N candidates are returned (billed as N × output_tokens)

1

Random seed (Beta) **Notes**: - When specified, the model attempts deterministic sampling - Same seed + same other parameters → same output (not guaranteed 100%)

0

Unique identifier representing the end user **Notes**: - Helps the platform monitor and detect abuse - A hashed user ID is recommended

"string"