Overview
DeepSeek is an AI assistant and open-weight model family from Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., whose parent company is the quantitative firm High-Flyer. There are three ways to use it: free chat on chat.deepseek.com or the official app; a usage-based API; and downloaded model weights that you can run on your own hardware.

The current official online models are the V4 series: `deepseek-flash` (DeepSeek-V4.1-Flash, updated on the official site to V4.1-Flash-1017, with native image understanding) and `deepseek-v4-pro` (DeepSeek-V4-Pro-0813). Both offer a 1 million-token context window and up to 384K tokens of output, with thinking and non-thinking modes. The distinctive combination is strong reasoning and coding, publicly released weights you can self-host, and API prices well below most leading closed models.
Official addresses all sit on the `deepseek.com` domain:
- Homepage: https://www.deepseek.com/
- Web chat: https://chat.deepseek.com/
- App and client downloads: https://www.deepseek.com/download/ (official mobile apps, plus the DeepSeek Harness desktop client)
- API platform: https://platform.deepseek.com/
- API docs and pricing: https://api-docs.deepseek.com/
> Features, model names, and prices change. The notes below follow the official site and the API pricing page, checked on 2026-11-11 (Asia/Shanghai). Re-check the official pages before you publish or budget.
Key Uses
- Everyday Q&A and web search: Ask in natural language. Turn on web search when you need fresh facts, and check the cited sources instead of trusting the summary alone.
- Deeper reasoning on hard problems: Use thinking mode for math, STEM, and multi-step logic. Leave it off for simple tasks. The API can set thinking effort to low, high, or max.
- Coding and debugging: Write, explain, refactor, and troubleshoot across languages. It can also back agents such as Claude Code, Codex, and Cursor.
- Long documents and large codebases: A 1 million-token context can hold a long report or a sizable repository. Load the relevant material, not every file you have.
- Image understanding: Send screenshots, charts, or photos to V4.1-Flash. V4-Pro does not support vision, so image tasks must use Flash.
- Writing and text work: Drafting, outlines, translation, editing, and summaries. Technical and explanatory prose in Chinese and English is solid; marketing copy and fiction are usually not a strength.
- Agents and tool calling: Function calling, structured JSON output, and the Responses API are supported. The API offers both OpenAI and Anthropic formats, so many existing clients only need a new base URL and model name.
- Private deployment: Weights are published, so regulated teams can run the model inside their own network. Follow the license in the official repository; the main cost is hardware and operations.
Who It Is For
- Developers and engineering teams who want to point existing OpenAI- or Anthropic-style code at a lower-cost model, or who are building coding agents.
- Cost-sensitive projects that run a lot of classification, extraction, or summarization.
- Teams that cannot send prompts off their own network and want to evaluate self-hosting.
- Students and researchers who can use strong math and reasoning ability for free in the web app.
- Legal, analysis, and editorial work that needs questions over a long document in one pass.
It is a weaker fit if you mainly need image or video generation, live voice, or deep product integrations with overseas office suites. In those cases, compare ChatGPT or self-host and handle compliance yourself.
How to use
- Use the web app or official mobile app for chat; use the API to integrate; download weights only if you need offline or private deployment. For long jobs that must read and write local files, see DeepSeek Harness.
- Register from www.deepseek.com. One account usually covers the web app, mobile app, and API platform.
- Use non-thinking mode for lookup, translation, and drafting; switch on thinking mode for math, logic, and multi-step work.
- Upload long files rather than pasting an entire document. On the web app, web search and file upload are often mutually exclusive, so split the work into two steps if you need both.
- Use Flash for images. Do not send pictures to Pro.
- Turn on web search when the answer depends on current information, and open the cited sources.
- For API access, set the base URL to `https://api.deepseek.com` (or `https://api.deepseek.com/anthropic` for the Anthropic format) and the model name to `deepseek-flash` or `deepseek-v4-pro`. Older names such as `deepseek-chat` and `deepseek-reasoner` have been retired.
- Choose the model by task: Flash is faster and cheaper, and it is the only vision option; Pro is for harder reasoning when quality matters more than unit cost.
- Keep the system prompt stable and put reusable context first so cache hits stay high; schedule batch jobs in off-peak windows.
- Review code, facts, and figures before you ship.
Pricing
Do not collapse this into a single “it is completely free” line. There are three paths:
- Web and official app chat: currently free. Thinking mode also does not require a consumer subscription. Unlike ChatGPT Plus, there is no paid consumer chat tier. Limits are more about busy-hour queues than a “messages used up, please pay” wall. If a page demands payment before you can chat, it is probably not the official product.
- API: billed by usage. Charges come from your platform balance, in US dollars, with separate rates for cache hits vs misses and peak vs off-peak. According to the official pricing page (confirm the live table):
- DeepSeek-V4.1-Flash, per million tokens: input cache hit $0.003 off-peak / $0.006 peak; input cache miss $0.15 off-peak / $0.30 peak; output $0.60 off-peak / $1.20 peak.
- DeepSeek-V4-Pro-0813, per million tokens: input cache hit $0.022 off-peak / $0.044 peak; input cache miss $0.66 off-peak / $1.32 peak; output $1.98 off-peak / $3.96 peak.
- Search and Fetch: $0.75 per 1,000 successful requests, billed separately from tokens.
- Self-hosted open weights: no model license fee, but you pay for hardware and operations. Full-size models need serious compute. A consumer PC usually runs a quantized or distilled variant, which is not the same as the official hosted service.
Off-peak prices are half of peak prices. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays — that is 09:00–12:00 and 14:00–18:00 in Beijing time (Asia/Shanghai, UTC+8). Weekends and Chinese public holidays are off-peak all day. Flash cache-hit input is much cheaper than a cache miss, so a stable system prompt at the start of the request can cut costs at scale.
Tips
- Leave thinking mode off for everyday tasks; reserve it for problems that need a real chain of steps.
- Put repeated context at the front of the prompt to improve cache hits.
- Run batch jobs off-peak (outside weekday Beijing peak hours, and on weekends and Chinese public holidays).
- Provide the surrounding module or full document, not a single function or excerpt.
- Say what the output is for (production code, a one-off script, or an explanation for a colleague).
- Configure model names in one place, so retired aliases do not linger in code.
Limitations
- Thinking mode is slower, and it costs more on the API.
- V4-Pro cannot read images.
- A 1 million-token window does not mean every section is remembered equally well, and cost still rises with input size.
- Peak hours are more expensive and more congested; you may hit rate limits or busy errors.
- Official web and API services are operated in China, and prompts are subject to local data and content rules. If that is not acceptable, evaluate self-hosting.
- English creative writing and marketing copy are not a strength.
- The official chat apps are not open-source software. What is released are the published model weights (and some related code). Follow the license in the official repository.
Frequently Asked Questions
What is DeepSeek, and which company is behind it?
DeepSeek is the AI assistant and model brand of Hangzhou DeepSeek. The parent company is High-Flyer. The product includes free web and app chat, a developer API, and downloadable open weights.
What is the official DeepSeek website?
The homepage is https://www.deepseek.com/, web chat is https://chat.deepseek.com/, downloads are at https://www.deepseek.com/download/, and the API platform is https://platform.deepseek.com/. Stay on the `deepseek.com` domain.
Is DeepSeek free?
Web and official app chat are currently free. The API is usage-based. Self-hosting the open weights has no model license fee, but you pay for hardware and operations.
Is DeepSeek open source? Can I use it commercially or self-host it?
Published model weights can be downloaded, run in your own environment, and used for lawful commercial purposes, subject to the official repository license. The web app and official apps are hosted services, not open-source software.
How should I choose between V4.1-Flash and V4-Pro?
Flash is faster and cheaper, and it is the only option with image understanding. Pro is stronger on hard reasoning and agent tasks. Default to Flash, and move a job to Pro only when quality is clearly falling short.
Is DeepSeek or ChatGPT the better choice?
For text, math, code, and long documents, DeepSeek is often enough, with free chat, a cheaper API, and weights you can self-host. If you need image generation, more mature voice features, or a broader assistant product, ChatGPT is still wider. Test both on your own tasks.
How do I download DeepSeek? Is there a desktop app?
For phones, use the official download page or the system app store, and confirm the developer is DeepSeek. On a computer, the simplest path is the web app in a browser. For a desktop client that can work with local files and long-running tasks, use official DeepSeek Harness, not a third-party “PC edition.” Downloading model weights is for local deployment, which is not the same as installing a chat app.
Can my existing OpenAI code talk to the DeepSeek API?
In most cases you change two things: set the base URL to `https://api.deepseek.com` and the model name to `deepseek-flash` or `deepseek-v4-pro`. An Anthropic-format endpoint is also available. Compatibility is not complete identity, so test with real requests before you go live.

