DeepSeek Open-Sources Official Frontend & Full-Capability API with Chat History Migration

DeepSeek Unifies App Capabilities into API and Open-Sources Frontend

DeepSeek has released a comprehensive API ( deepseek-flash, V4.1-Flash) that packages web search, URL reading, file parsing (PDF/Word), OCR, and voice interaction into the Messages, Responses, and Chat Completions interfaces. Simultaneously, the company open-sourced its official frontend chat-demo under MIT, a 1:1 replica of the web/app UI, and added a one-click import tool for migrating chat history from the official DeepSeek app.

Previous Developer Pain Points

Before this release, frontend or Node.js developers building a chat experience comparable to the official app had to stitch together multiple services: a third-party search API for web search, a headless browser or crawler for URL content extraction, separate parsers for PDF/Word uploads, and a speech-to-text service for voice input. The model API alone covered only the core LLM capability.

New Unified API: deepseek-flash (V4.1-Flash)

All peripheral capabilities are now built into the model API. A community meme illustrated the shift: a developer passes a single parameter and receives a full tray of building blocks — SEARCH, FILES, OCR, TTS — with the caption "take the whole app."

Under the /anthropic/v1 compatible protocol, the request structure is minimal. The TypeScript type for CreateMessageRequest shows the available switches:

export type CreateMessageRequest = {
  model: 'deepseek-flash' | 'deepseek-v4-pro'
  system: string
  messages: MessageParam[]
  thinking: { type: 'enabled' | 'disabled' }
  tools: (
    | { type: 'deepseek_web_search'; name: 'deepseek_web_search' }
    | { type: 'deepseek_read_url'; name: 'deepseek_read_url' }
  )[]
  safety_check: { level: 'high' }
  title?: { enabled: boolean }
}

Declaring deepseek_web_search in tools enables automatic web and image search with citation markers in non-thinking mode; in thinking mode the model interleaves search with multi-step reasoning. Declaring deepseek_read_url lets the service fetch and parse links from user messages. File uploads use /anthropic/v1/files with a rich_layout_recognition purpose flag to trigger server-side layout parsing for PDFs and Word documents.

Each capability is also exposed as a standalone endpoint: search and web fetch cost 5 CNY per 1,000 calls; voice recognition is 0.39 CNY per hour after a limited-time 50% discount.

Open-Source Frontend: chat-demo

The chat-demo repository (https://github.com/deepseek-ai/chat-demo) and online demo (https://cdn.deepseek.com/chat-demo) provide a ready-made template with zero backend code. The stack is modern and minimal:

React 18, Zustand, TypeScript 7.0

Rspack 2.0 with built-in SWC and LightningCSS (no Webpack, Babel, or ESLint)

Oxc ecosystem tools: oxlint for linting, oxfmt for formatting

After downloading the release build, developers can double-click index.html to run locally under the file:// protocol, enter an API key, and have a fully functional private client. All sessions and attachments are stored in the browser's IndexedDB.

One-Click Chat History Migration

The frontend includes an import feature compatible with DeepSeek's official deepseek_data export zip. Using fflate for decompression, it skips user.json (account info) and writes message branch trees, thinking durations, original images, and even previously fetched webpage signatures into the local database, allowing the server to restore context for those URLs.

Philosophy: Personal Super Apps

Project lead Tianyi Cui (崔添翼) explained the strategy using a smartphone OS analogy: the OS ships a basic calculator and alarm, but the vendor wants an ecosystem of third-party apps that surpass the built-ins — even replacing them. In DeepSeek Harness (the plugin system), every built-in feature is replaceable. Cui's latest statement underscores the vision: the next super app isn't one app for everyone; it's the ability for each person to craft a personal super app tailored to their own workflow. By making every low-level wheel available at minimal cost and ensuring historical data is portable, DeepSeek hands developers the raw materials to build exactly that.

References

chat-demo repository: https://github.com/deepseek-ai/chat-demo

Online demo: https://cdn.deepseek.com/chat-demo