Complete tutorial on using DeepSeek R1, getting started with the 2026 Chinese large model from scratch

DeepSeek is a Chinese AI large language model that burst onto the scene in early 2025. After the launch of its DeepSeek R1 reasoning version, three major advantages, open-source availability, strong reasoning, and rock-bottom API pricing, helped it surpass 150 million users within a single year, splitting the Chinese market evenly with ChatGPT. The latest version as of May 2026, DeepSeek R1 0428, scored 97.3 on the MATH-500 math competition benchmark, beating both Claude Opus 4.7 and GPT-4o.

This article walks through the complete workflow for using the latest May 2026 version of DeepSeek R1, suitable for everyone from total beginners to users who want to dig deep into the API. It covers four tiers, registration, basic web usage, the mobile app, API integration, and local deployment, plus 12 hands-on tested scenarios.

What Is DeepSeek, and How Is It Different from ChatGPT

DeepSeek is the large-model brand launched by the Hangzhou-based company DeepSeek (Deep Seek AI). It drew industry attention when it released V2 in May 2024, but it truly broke into the mainstream after the R1 reasoning model launched in January 2025. There are three key differences compared with ChatGPT.

The first is Chinese-language capability. Chinese accounts for over 30% of DeepSeek's training data, far higher than ChatGPT's 8% to 12%. Its accuracy is noticeably higher on Chinese writing, classical Chinese translation, and questions about Chinese cultural knowledge. For example, when you ask both to write a letter in classical Chinese, DeepSeek's sentence structures and literary allusions feel more authentic.

The second is mathematical and coding reasoning. The R1 model uses reinforcement learning to train a dedicated reasoning chain, and it can produce complete step-by-step solutions to competition-level math problems. Its first-attempt pass rate on LeetCode Hard problems is 75%, 7 percentage points higher than GPT-4o's 68%.

The third is price. With the DeepSeek API, 1 yuan buys you 2 million input tokens, and 1 yuan buys 800,000 output tokens. Compared with GPT-4o's 300,000 tokens per US dollar, that is more than 30 times cheaper. The same project budget can process 30 times the volume.

Web Registration and First-Time Use

Open chat.deepseek.com and click Sign Up. Registration supports three methods: phone number, email, and Google. Chinese users find phone-number registration most convenient, while overseas users can use one-click Google login. Once registration is complete, you are automatically taken to the chat page with no extra verification.

The left side of the page shows your conversation history, and the right side is the chat box. At the top there are two buttons, DeepThink and Search. DeepThink turns on R1 reasoning mode and is off by default; you don't need it for casual chat or simple Q&A, but once enabled the model first displays its thinking process before giving an answer. Search enables web access, after which answers will cite the latest web information, which suits questions about the latest developments in 2026.

For your first time, it's best to click New Chat to the right of the input box to start a clean conversation. Each conversation can hold up to 120,000 characters, roughly 60,000 Chinese characters, after which it automatically starts a new session.

How to Use DeepThink Reasoning Mode

DeepThink is R1's killer feature. Once enabled, the model first displays its reasoning process in a thinking block, and only then gives the final answer. This reasoning process is usually 500 to 5,000 words depending on the complexity of the question.

There are a few categories of scenarios well suited to DeepThink. The first is math problems, especially middle-school, high-school, and competition level. The second is coding problems that require a complete line of thought from algorithm to code. The third is logical reasoning problems such as logic paradoxes, Sudoku, and brain teasers. The fourth is business analysis, strategy questions that require multi-step deduction.

There are also scenarios where DeepThink is not appropriate. For tasks that don't need long reasoning, such as chatting, writing poetry, or translation, turning it on actually makes things slower. For short Q&A like checking the weather or converting units, turning it on wastes time. Each round of thinking consumes about 10 to 30 seconds, which is not worthwhile for scenarios with high real-time requirements.

In one hands-on test using a final, hardest question from the gaokao college-entrance exam, the V3 model answered incorrectly with DeepThink off. With DeepThink enabled, R1 thought for 47 seconds, output a 2,300-word reasoning process, and arrived at the correct final answer. This is why R1 is so widely used in China's education and tutoring sector.

12 High-Frequency Practical Scenarios

First, coding assistance. Ask R1 to write a Python web-scraping script, paste in the requirements and sample output, and it returns complete code with comments in 30 seconds. The error rate is half that of asking V3.

Second, Chinese polishing. Paste in a stiff passage that was translated from English, ask for natural Chinese, and the result beats Google Translate and DeepL.

Third, paper summarization. Paste in the full text of a PDF and ask for a summary within 500 characters, with accuracy above 90%. It's slightly faster than asking ChatGPT to summarize a document of the same length.

Fourth, resume optimization. Paste in your resume and say "rewrite it for an application to a ByteDance product manager role," and it automatically adjusts keywords and phrasing.

Fifth, email writing. "Help me write a reply to a client politely declining a meeting invitation, courteous but not subservient," and the output is ready to send.

Sixth, legal consultation. Ask "the rental contract doesn't mention it, but my landlord says the deposit won't be refunded, what should I do," and it gives you the legal basis and the steps to take.

Seventh, writing SQL. Describe the table structure and requirements, and the generated SQL not only runs but also comments on what each line does.

Eighth, math tutoring. The most common use for parents helping kids with homework. Turn on DeepThink and have the model solve problems step by step.

Ninth, creative copywriting. For Xiaohongshu-style, Douyin-style, and WeChat official-account-style titles and copy, DeepSeek fits the feel of Chinese social media better than GPT-4o.

Tenth, code review. Paste in a chunk of code and ask "what can be improved," and it points out three categories: bugs, performance issues, and style issues.

Eleventh, business planning. Turn on DeepThink and have the model analyze "can a coffee shop with 20,000 yuan in startup capital be viable," and it produces market research, cost estimates, risk analysis, and a 3-year financial model.

Twelfth, multi-round brainstorming. Keep probing an open-ended question, and the model proposes new ideas from different angles. It's 5 times more useful than a single-shot answer.

Getting Started with the DeepSeek Mobile App

The official DeepSeek app supports both iOS and Android. Search for DeepSeek in the App Store or an Android app store and download it; it's about 80 MB. Registration uses the same account as the web version, and it automatically syncs your conversation history.

The mobile-specific features are voice input and photo-based problem solving. For voice input, hold down the microphone button and speak, with support for mixed Chinese and English and transcription accuracy above 95%. For photo-based problem solving, point the camera at a textbook problem and take a picture, and it automatically recognizes the question and solves it, which suits students.

Compared with the web version, the app lacks an API entry point and model-version switching. If you need those two features you still have to go back to the desktop. The chat experience is identical, and the response speed is even faster than the web because the app uses a long-lived WebSocket connection.

The Complete API Integration Workflow

Go to platform.deepseek.com, complete real-name verification, and top up starting from 10 yuan. Click API Keys to generate a new key and save it carefully, as it is shown only once.

For a code example, just call it with the Python OpenAI SDK, because the DeepSeek API is fully compatible with the OpenAI format. Import the openai library, set base_url to https://api.deepseek.com, and fill in the api_key you just generated. The model name deepseek-chat is the V3 default version, and deepseek-reasoner is the R1 reasoning version.

In one hands-on test of a common script, batch-translating 1,000 English news summaries into Chinese, the V3 model took 6 minutes at a cost of 0.8 yuan. The same volume with GPT-4o took 12 minutes at a cost of 28 yuan. For this kind of repetitive, high-volume task, DeepSeek's value for money is unbeatable.

The API concurrency limit is 200 calls per minute, with a maximum single output of 8K tokens. If you're doing large batch tasks, it's advisable to add a 0.3-second sleep interval to avoid triggering rate limiting.

Locally Deploying the DeepSeek R1 Distilled Versions

Users who don't want to use cloud services can run it locally. But the full R1 is a 671B-parameter model with weights of 1.3TB, which an ordinary computer cannot run. The company has released 6 distilled versions, from 1.5B to 70B, suitable for different hardware.

The most suitable for personal use are the 7B and 14B versions. The 7B takes 4GB of VRAM or 8GB of RAM. An M1 MacBook Air with 16GB can run it smoothly at 12 tokens per second. The 14B takes 8GB of VRAM, and an RTX 3060 12GB card can run it at 20 tokens per second.

The simplest way to deploy is Ollama. On Mac, install it with Homebrew; on Windows, download the .exe. Once installed, run ollama run deepseek-r1:7b in the terminal to automatically download the model and start it. The first download is about 4GB and takes 5 to 10 minutes. Once it finishes downloading, a chat dialog pops up and you can start chatting.

The advantages of local deployment are data privacy and offline availability. Client privacy, contracts, and code won't be uploaded to a server. The downside is that the smaller model is weaker; the 7B distilled version is roughly equivalent to GPT-3.5 and not as capable as the full cloud R1.

Pricing and Quota Comparison

For the cloud API, the V3 model costs 0.5 yuan per million input tokens and 8 yuan per million output tokens, while the R1 reasoning model costs 1 yuan per million input tokens and 16 yuan per million output tokens.

By comparison, OpenAI GPT-4o costs 18 yuan per million input tokens and 60 yuan per million output tokens, and Anthropic Claude Opus 4.7 costs 110 yuan per million input tokens and 540 yuan per million output tokens. DeepSeek's pricing is 1/30 that of OpenAI and 1/100 that of Anthropic.

Top up 100 yuan and get 5% extra; top up 1,000 yuan and get 10% extra. New users get a 10-yuan credit on registration, enough for an ordinary user for a month. Small and medium developers can easily run on 100 yuan for half a year.

Data Security and Compliance

On data privacy, DeepSeek does not use user conversation data to train subsequent models. Conversations on the web version are stored for 30 days by default, and API calls do not record any conversation content. This is stricter than ChatGPT's early days, because OpenAI used user data for training back in 2023.

On compliance, DeepSeek has an ICP filing within China, with data stored on domestic servers. Enterprise customers that need data to leave the country can choose overseas nodes; deepseek-r1-instruct has public mirrors on Hugging Face and ModelScope that can be used commercially.

On international export controls, the United States briefly banned DeepSeek in January 2025 before lifting the ban. As of May 2026, the website and API can still be accessed normally in the US, but the iOS App Store has removed it in the US region.

Troubleshooting Common Usage Issues

First, lag. R1 reasoning mode occasionally hangs for a minute with no response. It's advisable to refresh the page and resend; in DeepThink mode, 90% of delays are within a minute, which is normal.

Second, truncated output. The model's maximum output is about 8K tokens, i.e., 4,000 Chinese characters. If you ask it to write a 10,000-character paper, it will stop halfway. The solution is to have the model continue in multiple passes.

Third, network problems. Access from mainland China is occasionally blocked; it's advisable to use China Telecom or China Unicom broadband, as accessing overseas nodes over a mobile network has high latency. Overseas users accessing domestic nodes face latency above 200 milliseconds and should switch to the Hugging Face mirror.

Fourth, mixed Chinese and English. When you ask the model to answer in English, it occasionally mixes in Chinese. Adding "Please respond in pure English only" to the system prompt fixes this.

Fifth, hallucinations. While reasoning, R1 occasionally fabricates API documentation and function names. For specific technical details, always cross-verify and never fully trust it.

Who DeepSeek Is Right For

Students are the best fit. Photo-based problem solving, math tutoring, paper summarization, and English translation, four things all handled in one app. At zero cost per month, the web version is enough.

Chinese content creators are the next best fit. Xiaohongshu bloggers, WeChat official-account writers, and self-media operators find DeepSeek's copywriting more down-to-earth than ChatGPT's. Drafting speed is 5 times faster than writing by hand.

Small and medium developers are an excellent fit. The API pricing lets small projects that couldn't afford OpenAI switch to something they can afford. For applications like batch translation, customer-service bots, and coding assistants, using DeepSeek directly offers the best value for money.

It's less suitable for non-Chinese scenarios. For an English product aimed at overseas users, the model is less natural in English than GPT-4o. For minor languages beyond Chinese, such as Japanese, Korean, and German, DeepSeek's output quality is also a notch below GPT's.

Frequently Asked Questions

Is DeepSeek Fully Free, and Is That Enough?

The web version is completely free with no limit on the number of calls; you can chat dozens or even over a hundred rounds a day. R1 reasoning mode has about a one-minute wait during peak times but is not charged. The web version is enough for 99% of ordinary users, including students, content creators, and product managers. You only need to pay when you require API integration, batch automation, or long-context scenarios above 120,000 characters. Topping up 10 yuan to the API lasts 2 to 3 months, so the barrier is essentially nil.

How Do I Choose Between R1 and V3?

V3 is the default model, suited to ordinary conversation, writing, and translation, with a fast response of 1 to 3 seconds and API output at 8 yuan per million tokens. R1 is the reasoning model, suited to math, programming, and logical analysis, with a slow response of 30 seconds to 2 minutes and API output at 16 yuan per million tokens, twice that of V3. The general rule is to use V3 for 80% of tasks, and R1 for the 20% of tasks that involve "thinking through how to do it."

Is a Locally Deployed 7B Model Enough for Daily Use?

The 7B distilled version is roughly equivalent to GPT-3.5; it can handle email writing, polishing, and translation, but it is noticeably worse than the cloud for creative writing, complex programming, and deep analysis. It's enough purely for offline scenarios, such as use on a business trip on a plane. For serious daily use, it's advisable to start at 14B or just use the cloud. The local 70B approaches the capability of the cloud R1 but needs 48GB of VRAM, runnable only on an RTX 5090 or A100, which is poor value for money.

Will DeepSeek's Output Be Used to Train New Models?

The official documentation clearly states that API call data has zero retention and is not used for training. Web conversations are stored for 30 days by default and storage can be turned off in settings. But sensitive data such as business contracts, personal ID numbers, and bank card numbers should still not be entered into any AI service. Local deployment is the only guarantee of zero data leaving your device.

Why Is DeepSeek So Much Cheaper Than ChatGPT?

Three reasons. First, the Chinese team's labor and electricity costs are 50% lower than Silicon Valley's. Second, the MoE mixture-of-experts architecture activates only some parameters during inference, so the compute overhead is small. Third, a scale-driven revenue strategy: DeepSeek doesn't chase short-term profit but rather seeks to capture market share. Prices are expected to rise 30% to 50% by 2027 but will still remain significantly below OpenAI's.