DeepSeek API Official References

Official References

Primary source: DeepSeek API Docs.

Quick start

• Your First API Call: base URL, API key, initial models, and chat example.
• Models & Pricing: available models, context, maximum output, features, and deprecations.
• Token & Token Usage: token definition and API-reported usage.
• Rate Limit & Isolation: limits, isolation, and user_id usage.
• Error Codes: expected HTTP errors and recommended handling.

API guides

• Thinking Mode: thinking, reasoning_effort, reasoning_content, and tool-call rules.
• Multi-round Conversation: message concatenation across turns.
• FIM Completion Beta: completion with prompt and suffix using the beta base URL.
• JSON Output: response_format: { type: "json_object" } and prompt requirements.
• Tool Calls: tool format, tool messages, strict mode, and supported schemas.
• Context Caching: context cache and cache hit/miss tokens.

API reference

• API Reference: Bearer authentication and endpoint map.
• Create Chat Completion: full /chat/completions contract.
• Create FIM Completion Beta: full completion/FIM contract.
• List Models: available models endpoint.

Rules that affect the code

• The main chat endpoint is POST /chat/completions.
• stream: true returns SSE events and ends with data: [DONE].
• reasoning_content appears separately from content in thinking mode.
• In thinking mode, temperature and top_p have no effect according to the official guide.
• Tool calls use tools, tool_choice, tool messages, and tool_call_id.
• Tool arguments arrive as a JSON string and must be validated before execution.
• JSON output requires response_format and prompt instructions asking for JSON.
• FIM beta uses https://api.deepseek.com/beta and a 4K token maximum.