DeepSeek Guide
Type "DeepSeek" into a search box and the results page turns into a hall of mirrors. There's the real site, an app-store listing, a couple of lookalike domains — extra hyphen, odd suffix — promising "free chat," and an explainer article still reviewing a model from early 2025. Three of those doors lead somewhere useful. The rest lead to knockoffs or stale advice. This guide exists to walk you past all of that: which doors are official, what the current models actually are, and how to put DeepSeek to work whether you want to chat, write, code, or build on the API.
What is DeepSeek?
DeepSeek is an AI research company based in Hangzhou, China — full name Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., registered on July 17, 2023. It was founded by Liang Wenfeng, a Zhejiang University graduate who also co-founded High-Flyer, the quantitative hedge fund that owns and funds the whole operation. Venture capitalists passed on the spin-off at first, per Wikipedia's company profile; the fund didn't. One founder, one backer, and a team of roughly 160 people as of 2025.
For context on how unusual that is: OpenAI reportedly spent over $100 million training GPT-4, while DeepSeek claimed about $6 million for the V3 generation that preceded its breakout. The budget debate never fully settled — the weights got published either way.
The company builds open-weight large language models — the model files are published for anyone to download and run, under a free MIT License for everything released since January 2025. Its breakout moment came fast. On January 20, 2025, DeepSeek released R1, a reasoning model that held its own against far more expensive rivals, alongside a free chatbot app. Within a week the app had passed ChatGPT as the most-downloaded free app on the US App Store, and Nvidia shed around 18% of its value in a single day — at the time, the largest one-day value loss any company in US stock market history had suffered.
That was the old chapter. The current one is the V4 family, released through 2026, and the rest of this guide keeps its footing there rather than relitigating 2025.
Official DeepSeek Websites & Apps
Start here, because this is where the hall of mirrors gets dangerous. The official site anchors a short list of real doors, and this section is your address guide to all of them:
| Entry | Address | What it's for |
|---|---|---|
| Main website | https://deepseek.com/ | Product hub, announcements, downloads |
| DeepSeek Web (chat) | https://chat.deepseek.com/ | Free chat interface in the browser |
| DeepSeek App | App Store & Google Play | Official iOS and Android app (via "Get App" on the main site) |
| API Platform | https://platform.deepseek.com/ | Create API keys, manage billing |
| API Docs | https://api-docs.deepseek.com/ | Quick starts, model reference, pricing tables |
| API Pricing | linked from the main site's footer | Current per-token prices |
| Service Status | linked from the main site's footer | Live outage and incident reports |
Now the warning shot. Within days of DeepSeek's app hitting number one, a wave of counterfeit domains surged — Wikipedia documents the registration spike starting January 26, 2025 and peaking two days later, built for phishing and typosquatting. The fakes follow a pattern: an extra hyphen in the name, an unusual suffix, a homepage that looks almost right. Think of a street where the official storefront is surrounded by knockoff signs using the same logo — inside those doors, your credentials and payment details are the product.
So treat the table above as the complete list of real addresses, and when in doubt, type deepseek.com yourself rather than following a search result or an ad. One more wrinkle worth knowing: even the "DeepSeek" trademark in the US was filed by an unrelated company 36 hours before DeepSeek's own application — brand confusion isn't just a fringe problem, it's baked into the landscape.
DeepSeek Models Today: The V4 Family
Most explainers you'll find are still describing the 2025 lineup. This guide skips that stale layer: here's what exists now. The V4 series arrived as a preview on April 24, 2026, with official releases following in the summer. Both models are MIT-licensed and share a one-million-token context window:
| Model (API name) | Role | Scale | Notes |
|---|---|---|---|
deepseek-v4-flash | Fast and cheap | 284B parameters | Updated to the 0731 build; non-reasoning, reasoning, and "Max" extended-reasoning modes |
deepseek-v4-pro | Flagship, deep thinking | 1.6T parameters | Updated to the 0813 build; stronger agent capabilities, Responses API and Codex integration |
deepseek-v4-flash-vision-exp | Experimental | — | Accepts image input; explicitly experimental, per the API docs |
A helpful mental model: economy and first class on the same plane. Flash gets you there quickly and cheaply; Pro burns more thinking time for harder problems. Neither is "better" — they're tickets for different trips. The date suffixes (0731, 0813) are version stamps: request the plain name and you always get the latest build, a convention the API docs spell out on their first page.
The road here, compressed into one table:
| Release | When | Why it mattered |
|---|---|---|
| DeepSeek-Coder | October 2023 | First model out the door, built for code |
| V2, V3 | 2024 | The efficiency architecture that made heads turn |
| R1 | January 2025 | Reasoning model + free app; App Store number one |
| V3.1, V3.2 | 2025 | Hybrid thinking modes, sparse attention refinements |
| V4 family | 2026 | Current generation — Flash, Pro, and the vision experiment |
Semiconductor makers like Huawei and Cambricon have already adopted V4, per Wikipedia — the ecosystem moved with it. There's also DeepSeek Harness, the company's open-source agent framework, currently in developer preview; worth knowing it exists, too early for most people to care.
What You Can Do with DeepSeek
Here's the useful reframing, and the reason this guide organizes by activity instead of by product: every entry point below is just a different on-ramp to the same highway. Same V4 models, same intelligence — what changes is how you drive onto it.
Chat and write
The simplest door: DeepSeek's web chat or the mobile app. Free to use, no API key, no billing. Draft emails, summarize documents, brainstorm, ask questions. For most people, this is all DeepSeek ever needs to be — a capable assistant that costs nothing.
Code with it
If you live in a terminal or an IDE, you can skip the chat window entirely. DeepSeek's API works as a backend model for Claude Code, GitHub Copilot, and OpenCode — officially supported, and as the API docs put it, no code required. You change a configuration, and your existing coding assistant starts running on DeepSeek's models.
Build with the API
For developers wiring DeepSeek into their own products, the platform speaks two dialects: an OpenAI-compatible format at https://api.deepseek.com and an Anthropic-compatible format at https://api.deepseek.com/anthropic. If your tool already talks to either, you point it at a new address and go. Requests can toggle extended thinking and reasoning effort, so one model serves both quick lookups and hard analysis.
Run it locally
Because the weights are open and MIT-licensed, you can download the models and run them on your own hardware, disconnected from DeepSeek's servers entirely. The distinction matters: the official chat and API are services — they run on the company's machines. The open weights are the engine itself. Installing that engine yourself is real work and real hardware, but it's the only path that keeps every byte on your own machines. What you can't do is use the official app offline; open weights and hosted service are different things, and conflating them is how misleading tutorials get written.
API usage is billed separately from the free chat — the early promotional rates ended in February 2025, so check the current pricing page rather than trusting numbers from old articles. Including, frankly, the ones still circulating.
DeepSeek FAQ
Is DeepSeek free?
The web chat and mobile app are free. The API is pay-per-use, priced per token. The open weights are free to download, though running them yourself requires serious hardware.
What's the difference between V4 and R1?
Different generations. R1 was the January 2025 reasoning model that made the company famous. V4, released through 2026, is the current family — Flash for speed, Pro for depth. Date suffixes in the names mark each refresh. The full story of that rise is in What is DeepSeek?.
Is DeepSeek safe to use?
For casual chat, standard caution applies. It's worth knowing the full picture: data practices drew regulatory attention in 2025, several governments have restricted the app on official devices, and a security research team found an exposed internal database in January 2025 that was quickly locked down. Sensitive corporate or personal data belongs elsewhere; Wikipedia's article on the company covers the record in detail.
Does DeepSeek have an official app?
Yes — iOS and Android, both linked from the "Get App" button on the official site. Install from the stores via that link, not from a download site claiming to offer the "official version."
DeepSeek or ChatGPT?
Different companies, different philosophies. DeepSeek's edge is open weights and a cheap, dual-compatible API; OpenAI's is breadth of ecosystem. The honest answer depends on what you're doing with it, and it deserves a fuller comparison than this guide can give in a FAQ line.
This DeepSeek guide anchors our collection — model deep-dives and setup walkthroughs will hang off it as they land.