DeepSeek

DeepSeek-V4 Models

DeepSeek's V4 model family includes two API tiers — V4-Flash for fast general-purpose inference and V4-Pro for maximum reasoning performance. Both run on a Mixture-of-Experts architecture with a 1M token context window and 384K max output. Models are accessible via the free web chat at chat.deepseek.com, iOS/Android apps, and the REST API. Context caching reduces repeated prefix costs by 90%.

Thinking Mode (Chain-of-Thought Reasoning)

Before outputting a final answer, DeepSeek's thinking mode produces a chain-of-thought reasoning trace (reasoning_content) to improve accuracy on complex tasks. Thinking mode is the default on V4-Pro and available on V4-Flash. Outputs run up to 384K tokens for extended reasoning chains.

Tool Calls (Function Calling)

DeepSeek's API allows the model to call external functions and tools to enhance its capabilities. Supported in both thinking and non-thinking modes, with a strict mode (Beta) that enforces exact JSON schema compliance on tool outputs.

JSON Output (Structured Output)

DeepSeek provides a JSON Output mode that guarantees the model returns valid, parseable JSON strings. Users set the response_format parameter to json_object and include a JSON example in their prompt to guide structured output for downstream parsing.

Context Caching (KV Cache on Disk)

The DeepSeek API automatically caches request prefixes on disk for all users at no extra configuration. Repeated prefix tokens in subsequent requests trigger a cache hit, reducing cost from $0.28 to $0.0028 per 1M input tokens — a 90% discount.

FIM Completion — Fill in the Middle (Beta)

Users can supply a code or text prefix and optional suffix, and the model fills in the middle content. Commonly used for code completion and content completion tasks, integrating natively with VS Code via the Continue plugin. Non-thinking mode only (V4 restriction). In V3.2, FIM was available in both modes; V4 restricts it to non-thinking mode.

Chat Prefix Completion (Beta)

Developers can provide an assistant message prefix and force the model to continue from exactly that point, enabling precise control over output format and style — for example, prefixing with a Python code block to guarantee code-only output.

Multi-round Conversation Support

The DeepSeek chat API supports stateful multi-turn conversations by allowing developers to concatenate full conversation history per request. The API is stateless by design, giving developers full control over context management across conversation turns.

OpenAI-Compatible API Format

The DeepSeek API is fully compatible with the OpenAI SDK and format, requiring only a base_url swap to https://api.deepseek.com. Existing OpenAI integrations can switch to DeepSeek models with minimal code changes.

Anthropic API Compatibility

DeepSeek's API supports the Anthropic API format, allowing teams using the Anthropic SDK or Claude Code to route requests to DeepSeek models. Supported fields include streaming, system prompts, tool use, thinking blocks, and temperature control.

DeepSeek Web Chat & Mobile App

DeepSeek offers a free web chat at chat.deepseek.com and a mobile app on iOS and Android. Features include web search, thinking mode (chain-of-thought), file upload with text extraction, and cross-platform chat history sync — with no ads or in-app purchases.

Open-Source Model Weights

DeepSeek releases model weights for DeepSeek-V4, DeepSeek-R1, DeepSeek-Coder V2, DeepSeek-VL, and others on Hugging Face under open licenses. This enables self-hosting, fine-tuning, and research use without API dependency.