DeepSeek Login and Download: What Works From the US

DeepSeek is free to use in a browser, and the desktop app is optional. The download is the easy part. The sign-in screen is where US users hit a wall, because the chat site and the developer platform do not offer the same login options.

I loaded every entry point on September 29, 2026 and wrote down what each one asks for. What follows is the account route, the fix for Chinese replies, and the price table with one timing detail that matters if you pay per token.

Last updated: September 29, 2026. Login screens are the kind of thing that changes without notice. The price table and the walkthrough are the durable part.

The four ways in

DeepSeek is not one product. It is a chat site, two mobile apps, a desktop app, plus an API. They do not all ask for the same things.

Three of the four are free, and the free one is the one you should start with. The exception is the API. The API is also the reason the account question further down actually matters.

I went through what that desktop app actually gives you in a separate post. Short version: it is a release candidate, and it is aimed at people who want an agent touching files. It is not a nicer chat window.

The mobile apps deserve one paragraph, because they are the option most people end up using and the least interesting to read about. They are free. They take the same account as the website. And they add exactly one thing the browser cannot give you, which is a home-screen icon. That is the whole upgrade.

So install one if you want. Just do not expect it to be a different product. It is the same model behind a smaller screen, and the reviews that describe it as a separate experience are describing the app shell.

The sign-in screen is where it gets specific

The chat site and the developer platform share one account. They do not share one login form. That is the first thing worth knowing.

People search for a DeepSeek Google login constantly. Neither screen offers one. Better to hear that now than after twenty minutes of hunting for a button that is not there.

Here is the judgement call I would make. If an API key is even a maybe for you, create the account with an email and a password. Sign in with Apple is the fastest route into the chat. The platform screen lists a phone number and a password instead, and then a WeChat QR code after that.

Setting a separate password today costs you half a minute. Doing it next month, with conversations and a topped-up balance already attached, is the annoying version.

I might be overweighting the account choice here. Apple sign-in works, it is faster, and DeepSeek has not published anything about linking a password later. But the platform screen I loaded did not list Apple, and I would rather spend thirty seconds than gamble on a support ticket in a language I do not read.

Why it keeps answering in Chinese

This is the most common complaint about DeepSeek. And it is not an account problem. Interface language and answer language are two different settings.

Changing the menus to English does not change how the model replies. Most advice online stops at “switch the site to English”, which fixes the buttons and nothing else.

There is no answer-language setting in the chat. There is no language parameter in the API either. The documented route is an instruction inside your message. DeepSeek’s own change log records language-consistency work in the V3.1-Terminus update, so the behaviour has been addressed at the model level. It has not been removed.

The trigger that catches people is a shift to short follow-ups. A long English thread is going fine. You paste a block of Chinese source text. You type “Translate this.” And the next three answers come back mixed. “Translate this” contains no language instruction, so the model picks one.

The official English homepage is at deepseek.com/en. Use it to reach the correct product links, then open the chat from there. And do not go looking for a separate English chat domain, because there is not one.

What it costs, and the one timing detail

The browser chat is free to use. The API is billed per token. The rates have a wrinkle that changes what you actually pay.

Off-peak is exactly half of peak. Peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. Chinese public holidays are excluded and count as off-peak.

Now put those two windows on a Chinese clock. They are 9am to noon, and 2pm to 6pm, Beijing time. The peak rate is a Chinese workday, billed by the hour.

That has a direct consequence for anyone in the United States. The expensive window lands at 9pm to midnight and 2am to 6am Eastern, or 6pm to 9pm and 11pm to 3am Pacific. If your jobs run during your own working day, you are already buying at the half-price rate. Weekends are off-peak in full.

So I would not build a scheduler around this. The saving is real, but it applies to hours most US users would not pick anyway.

What I would pay attention to is caching. A cache hit on the Flash model costs $0.003 per million input tokens against $0.15 for a miss. That is a fifty-fold gap. It dwarfs the peak discount. If your app resends the same long system prompt on every call and you are not seeing cache hits, fix that before you touch anything else.

Setting up an API key without guessing

The API is where DeepSeek stops being a website and starts being a line item. Four things have to line up, and the documentation assumes you already know all four.

The base URL wastes the most time, because DeepSeek publishes two of them. Two URLs. Pick one. The first speaks the OpenAI request format and the second speaks Anthropic’s. Match it to your client and leave the other alone.

Billing runs off a balance you top up in advance. When a granted balance and a paid balance both exist, the granted one is spent first. That ordering matters if you are tracking spend against a budget, because your card will not move until the free credit is gone.

And check the concurrency ceiling before you design anything. It is 2,500 for the Flash model and 500 for v4-pro. A fan-out workload will hit that difference long before it hits the price difference.

Where people get stuck

Most “DeepSeek not working” reports resolve into one of five things.

The retired-name row is the one that quietly costs people money. A client sending the old model string keeps working. So nothing forces you to notice that you are running a different model than the one written in your notes.

Both current models take a 1 million token context window and can return up to 384K tokens. Those numbers are large enough that context is rarely your constraint. Cost is.

Which one I would pick

Four entry points is three more than most people need. Here is how I would choose.

The combination I would avoid is signing in with Apple and then deciding you want an API key. It is fixable, and it is a bad first hour.

If you want the wider picture, the same three-door check across ten Chinese AI apps covers the field in one table. And if you are comparing models rather than products, the pricing gap against GPT-6 is the other half of this story. Both sit under the China Tech in America index, which is where I keep the rest of it.

What I would watch next is the desktop app. It is the only piece here shipped through a download host rather than a store, and the only piece still labelled as a candidate. Everything else on this page should still hold in a year.