DeepSeek API

Provides drop-in compatible REST API endpoints matching OpenAI's chat completion and embedding interfaces, allowing existing OpenAI client libraries (Python, Node.js, Go, etc.) to route requests to DeepSeek models without code changes. Implements request/response schema parity with OpenAI's API including streaming, function calling, and token counting, enabling zero-friction migration from OpenAI to DeepSeek infrastructure.

Exposes DeepSeek-R1, a reasoning-specialized model that performs explicit chain-of-thought computation before generating responses, using an internal reasoning token budget to decompose complex problems. The API returns both the reasoning trace (via special tokens or metadata) and the final answer, enabling applications to inspect the model's problem-solving process and validate correctness for high-stakes tasks.

Supports variable context windows (4K, 8K, 32K, 128K tokens depending on model) allowing applications to include more or less context based on requirements. The API accepts full conversation history and context, and applications can implement dynamic optimization strategies (summarization, retrieval-augmented generation, or sliding window) to stay within context limits while preserving relevant information.

Provides versioned API endpoints and model identifiers (e.g., deepseek-chat, deepseek-coder, deepseek-r1) with clear deprecation timelines, allowing applications to pin specific model versions and migrate gradually to newer versions. The API maintains backward compatibility for deprecated models during transition periods, and provides migration guides and performance comparisons to help teams evaluate upgrades.

Provides specialized code generation capabilities across 40+ programming languages (Python, JavaScript, Go, Rust, Java, C++, etc.) using DeepSeek-V3's training on diverse code repositories. The API accepts partial code, docstrings, or natural language descriptions and generates syntactically valid, contextually appropriate code completions. Supports both single-line completions and full function/class generation with awareness of language-specific idioms and frameworks.

Implements server-sent events (SSE) based streaming that delivers model outputs token-by-token in real-time, allowing clients to display partial results as they arrive rather than waiting for full completion. The API returns structured JSON events containing individual tokens, token probabilities, and cumulative token counts, enabling applications to implement progressive UI updates, early stopping, or dynamic prompt adjustment based on partial outputs.

Implements OpenAI-compatible function calling that allows models to request execution of external tools by generating structured JSON function calls matching predefined schemas. The API accepts a list of function definitions (name, description, parameters as JSON schema) and returns function call requests when the model determines a tool is needed, enabling agentic workflows where the model orchestrates multi-step tasks by calling external APIs, databases, or services.

Provides a batch processing endpoint that accepts multiple requests in JSONL format and processes them asynchronously at reduced rates (typically 50% discount vs on-demand pricing). The API queues batch jobs, processes them during off-peak hours, and returns results via webhook or polling, enabling cost-effective processing of large volumes of inference requests without real-time latency requirements.

Provides a dedicated token counting endpoint that accepts prompts and returns exact token counts for input and estimated output tokens, allowing applications to calculate costs before making requests. The endpoint uses the same tokenizer as the inference engine, ensuring accuracy for cost estimation and quota management. Supports counting tokens for chat messages, function definitions, and system prompts with language-specific tokenization rules.

Implements stateless conversation handling where clients manage conversation history by including full message arrays in each request, with the API maintaining no server-side session state. The API accepts a messages array (system, user, assistant messages in chronological order) and generates the next response while preserving context from previous turns. Supports conversation branching, message editing, and context window management through client-side logic.

Provides open-source versions of DeepSeek models (e.g., DeepSeek-7B, DeepSeek-33B) available on Hugging Face that can be self-hosted on private infrastructure using standard frameworks (vLLM, Ollama, llama.cpp, etc.). Enables organizations to run DeepSeek models on-premises with full control over data, latency, and costs, while maintaining compatibility with the same prompting and function-calling patterns as the API.

Provides a dedicated embedding endpoint that converts text into fixed-dimensional dense vectors (typically 1536 or 3072 dimensions) suitable for semantic search, clustering, and similarity comparison. The embeddings are trained on diverse text corpora and optimized for retrieval tasks, enabling applications to build vector databases, implement semantic search, or compute text similarity without training custom embedding models.