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

DeepSeek is a large Chinese AI model that came out in early 2025. After the DeepSeek R1 inference version was launched, it relied on the three major advantages of open source, strong inference capabilities, and low API prices. The number of users exceeded 150 million in one year, and it was evenly matched with ChatGPT in the Chinese market. The latest version of DeepSeek R1 0428 in May 2026 scored 97.3 points in the mathematics competition MATH-500, surpassing Claude Opus 4.7 and GPT-4o.

This article summarizes the complete usage process of the latest version of DeepSeek R1 in May 2026, which is suitable for users who want to use the API in depth from scratch. It includes four levels: registration, basic web gameplay, mobile app, API access, local deployment, and 12 actual test scenarios.

What is DeepSeek and what is the difference between it and ChatGPT

DeepSeek is a large model brand launched by Hangzhou DeepSeek Company. The V2 version attracted industry attention when it was released in May 2024, and it really hit the scene after the release of the R1 inference model in January 2025. There are three differences from ChatGPT.

The first is Chinese language ability. The proportion of DeepSeek training data in Chinese accounts for more than 30%, which is much higher than the 8% to 12% of ChatGPT. The accuracy of answering questions on Chinese writing, classical Chinese translation, and general knowledge about Chinese culture is significantly higher. For example, if the two are asked to write a letter in classical Chinese, the sentence patterns and allusions output by DeepSeek are more authentic.

The second is mathematical and programming reasoning. The R1 model uses reinforcement learning to train a specialized reasoning chain, and can provide complete step-by-step answers to mathematics competition-level questions. LeetCode's first-time pass rate for difficult questions is 75%, which is 7 percentage points higher than GPT-4o's 68%.

The third is price. DeepSeek API inputs 1 yuan to buy 2 million tokens, and outputs 1 yuan 800,000 tokens. Compared with GPT-4o’s 1 USD and 300,000 tokens, it is more than 30 times cheaper. The same project budget can run 30 times the volume.

Web registration and first use

Open chat.deepseek.com and click Sign Up. Registration supports three methods: mobile phone number, email, and Google. It is most convenient for Chinese users to use their mobile phone number, and overseas users can use Google to log in with one click. After registration is completed, you will automatically jump to the conversation page without additional verification.

The left side of the page is the conversation history and the right side is the chat box. There are two buttons at the top DeepThink and Search. When DeepThink is turned on, the R1 reasoning mode is turned off by default. If it is not needed for chatting or simple Q&A, it will first display the thinking process and then give the answer. Search is an online search that will answer questions that quote the latest web page information and are suitable for checking the latest developments in 2026.

For first-time users, it is recommended to click New Chat on the right side of the input box to start a clean conversation. The maximum length of each conversation is 120,000 characters, approximately 60,000 Chinese characters. If the limit exceeds 60,000 Chinese characters, a new conversation will be automatically opened.

How to use DeepThink inference mode

DeepThink is the R1’s killer feature. After being turned on, the model will first display the reasoning process in the thinking block before giving the final answer. This reasoning process is usually 500 to 5000 words depending on the complexity of the problem.

There are several scenarios suitable for opening DeepThink. The first is math questions, especially at the junior high and high school competition levels. The second is that programming questions require a complete idea from algorithm to code. The third is logical reasoning questions such as logical paradox Sudoku brain teasers. The fourth is strategic issues that require multi-step deduction in business analysis.

Not suitable for DeepThink scenarios. Tasks such as chatting, writing poems, and translating that do not require long reasoning are slower. Short questions and answers, such as checking weather conversion units, are a waste of time. Each thought will consume about 10 to 30 seconds, which is not cost-effective in scenarios with high real-time requirements.

A practical test of a final question in the college entrance examination. Turn off DeepThink Answering Errors with V3 Models. Turn on DeepThink R1 and think for 47 seconds to output 2300 words of reasoning process and the final answer is correct. This is the reason why R1 is widely used in China’s education and training field.

12 high-frequency practical scenarios

First programming aid. Let R1 write a Python crawler script, post the sample output of the requirements post, and return the complete code with comments in 30 seconds. Half the write error rate than letting V3 do.

Second Chinese polish. Paste a piece of stiff text translated from English to it and ask it to be changed into natural Chinese. The effect is better than Google Translate and DeepL.

Abstract of the third paper. The full text of the pasted PDF requires a summary within 500 words, with an accuracy rate of over 90%. Slightly faster than letting ChatGPT summarize documents of the same length.

Fourth resume optimization. Paste your resume and say "Change to the version for applying for the ByteDance product manager position", and the keywords and sentence patterns will be automatically adjusted.

Fifth, write an email. "Reply an email to the client declining the meeting invitation in a polite but not condescending tone." The output can be sent directly.

Sixth legal advice. If you ask "What should I do if the landlord says the deposit will not be refunded even though it is not written in the rental contract?", the legal basis and operational steps will be given.

Seventh, write SQL. Describe the table structure and requirements, and the generated SQL is not only executable but also annotates the role of each row.

8th Mathematics Tutoring. It is most commonly used by parents to help their children with homework. Open DeepThink and let the model solve the problem step by step.

Ninth creative copywriting. For title copywriting in Xiaohongshu style, Douyin style, and public account style, DeepSeek is more in line with the language of Chinese social media than GPT-4o.

10th Code Review. Post a piece of code and ask "What can be improved?" Three categories will be pointed out: bugs, performance issues, and style issues.

Eleventh Business Planning. Open DeepThink and let the model analyze "Can a coffee shop with a starting capital of 20,000 yuan be opened?" It will provide market research, cost estimation, risk analysis, and a three-year financial model.

Twelve rounds of brainstorming. By asking an open question repeatedly, the model will come up with new ideas from different angles. 5 times better than a single answer.

Get started with DeepSeek mobile app

DeepSeek official app supports iOS and Android. Search DeepSeek and download it from the App Store or Android App Store. The size is about 80 MB. If you register and use the same account on the web page, the conversation history will be automatically synchronized.

The special functions of the mobile terminal are voice input and taking pictures to solve problems. Voice input: Press and hold the microphone button to speak. It supports mixed Chinese and English, and the transcription accuracy is more than 95%. Take a picture to solve the problem. Take a picture of the textbook topic, automatically identify the question and answer it, which is suitable for students.

The App lacks API entry and model version switching compared to the web version. If you need these two functions, you still have to go back to the computer. The chat experience is exactly the same, and the response speed is even faster than the web page because the App uses WebSocket long connection.

API access complete process

Enter platform.deepseek.com and recharge starting from 10 yuan after real-name authentication. Click API Keys to generate a new key, save it and display it only once.

The code example can be called 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 api_key with the one you just generated. The model name deepseek-chat is the default version of V3, and deepseek-reasoner is the R1 reasoning version.

Test a common script. To batch translate 1,000 English news summaries into Chinese, the V3 model takes 6 minutes and costs 0.8 yuan. Using GPT-4o for the same amount of time takes 12 minutes and costs 28 yuan. DeepSeek is very cost-effective for such repetitive tasks.

The API call concurrency limit is 200 times per minute with a maximum single output of 8K tokens. If you do a large number of batch tasks, it is recommended to add a sleep interval of 0.3 seconds to avoid triggering current limiting.

Deploy DeepSeek R1 distilled version locally

Users who do not want to use cloud services can run it locally. But the complete R1 is 671B parameter IP and weighs 1.3TB, which cannot be run by ordinary computers. Officially released 6 distilled versions ranging from 1.5B to 70B suitable for different hardware.

The most suitable for personal use are the 7B and 14B versions. 7B takes up 4GB of video memory or 8GB of RAM. The M1 MacBook Air 16GB can run smoothly and output 12 tokens per second. 14B occupies 8GB of video memory. The RTX 3060 12GB card can run 20 tokens per second.

The easiest way to deploy is Ollama. Install with Homebrew on Mac, download .exe on Windows. After the installation is complete, run ollama run deepseek-r1:7b in the terminal to automatically download the model and start it. The first download of 4GB takes about 5 to 10 minutes. After downloading, a dialog box will pop up and you can chat.

The advantages of on-premises deployment are data privacy and offline availability. Customer privacy contract code will not be uploaded to the server. The disadvantage is that the model is small and the capabilities are weak. The capabilities of the 7B distilled version are approximately equivalent to GPT-3.5 and are not as good as the full version of Cloud R1.

Price and quota comparison

Cloud API Price The V3 model inputs 0.5 yuan per million tokens and outputs 8 yuan per million tokens. The R1 inference model inputs 1 yuan per million tokens and outputs 16 yuan per million tokens.

Compared with OpenAI GPT-4o, which inputs 18 yuan per million tokens and outputs 60 yuan per million tokens, Anthropic Claude Opus 4.7 inputs 110 yuan per million tokens and outputs 540 yuan per million tokens. DeepSeek’s price is 1/30 of OpenAI and 1/100 of Anthropic.

Recharge over 100 yuan and get 5%, recharge over 1,000 yuan and get 10%. The 10 yuan credit for new user registration is enough for ordinary users to use for one month. Small and medium-sized developers can run for half a year with 100 yuan.

Data security and compliance

In terms of data privacy, DeepSeek will not use user conversation data to train subsequent models. User conversations on the web are stored for 30 days by default. API calls do not record any conversation content. This is stricter than in the early days of ChatGPT because OpenAI was trained on user data in 2023.

At the compliance level, DeepSeek has ICP filing data stored in domestic servers in China. Enterprise customers who need to export data can choose overseas nodes. deepseek-r1-instruct has public images on huggingface and modelscope for commercial use.

In terms of international export controls, the United States banned DeepSeek in January 2025 and later lifted the ban. In May 2026, web pages and APIs can still be accessed normally in the United States, but the US version of the iOS App Store will be removed.

Frequently asked questions and demining

The first lag. R1 inference mode occasionally freezes for 1 minute without responding. It is recommended to refresh the page and resend. 90% of the delay in DeepThink mode is normal within 1 minute.

The second output is truncated. The maximum output of the model is about 8K tokens, which is 4000 Chinese characters. If asked to write a 10,000-word paper, it would stop halfway through. The solution is to continue writing the model multiple times.

Third network problem. It is recommended to use China Telecom or China Unicom broadband when accessing mainland China is occasionally blocked. Mobile network access to overseas nodes has a large delay. If overseas users have a delay of more than 200 milliseconds when accessing domestic nodes, it is recommended to switch to the huggingface mirror.

Fourth, Chinese and English are confused. When the model is asked to answer in English, Chinese and English will occasionally be mixed. Adding "Please respond in pure English only" to the system prompt can solve the problem.

The fifth illusion. R1 occasionally makes up API documentation and function names when reasoning. Specific technical details must be cross-verified and cannot be completely trusted.

Who is DeepSeek suitable for?

Best suited for student groups. One app can solve all four tasks including photo problem solving, math tutoring, thesis summary, and English translation. The web version costs 0 per month and is enough.

Secondly suitable for Chinese content creators. Xiaohongshu bloggers, public account authors, and self-media operators use DeepSeek to write copy more down-to-earth than ChatGPT. The output speed is 5 times faster than handwriting.

Very suitable for small and medium-sized developers. API prices have made OpenAI affordable for small and medium-sized projects. For applications such as batch translation, customer service bots, and code assistants, DeepSeek is the most cost-effective solution.

Not suitable for non-Chinese scenes. When making English products for overseas users, the model is not as natural as GPT-4o in the English field. For minor languages ​​other than Chinese, such as Japanese, Korean, and German, the output quality of DeepSeek is also somewhat worse than that of GPT.

FAQ

Is DeepSeek completely free enough?

The web page is completely free and has no limit on the number of calls. You can chat dozens or even hundreds of rounds every day. There will be a waiting time of about 1 minute during the peak period of R1 inference mode, but there is no charge. 99% of ordinary users use the web page, including students, content creators, and product managers. You only need to pay when you need API access, batch automation, and long context scenarios of more than 120,000 words. API can be used for 2 to 3 months by recharging 10 yuan, and there is basically no threshold.

How to choose between R1 and V3

V3 is the default model suitable for ordinary conversation writing and translation, with a fast response of 1 to 3 seconds, and an API output of 8 yuan per million tokens. R1 is an inference model suitable for mathematical programming logic analysis. The response is 30 seconds to 2 minutes slow. The API output is 16 yuan per million tokens, which is twice that of V3. The general rule is to use V3 for 80% of the tasks, and use R1 for the 20% of tasks that involve "figuring out how to do it."

Is the local deployment of the 7B model sufficient for daily use?

The capabilities of the 7B distilled version are approximately equivalent to GPT-3.5. It can be used to write emails and polish translations, but creative writing, complex programming, and in-depth analysis will be significantly worse than the cloud version. It is enough for offline scenarios such as business trips. For serious daily use, it is recommended to start from 14B or use the cloud directly. The local 70B is close to the cloud R1 capability but requires 48GB of video memory. Only RTX 5090 or A100 can run it, which is not cost-effective.

Will DeepSeek output be used to train new models?

The official documentation clearly says that API call data with zero retention will not be used for training. Web conversations are saved for 30 days by default, and storage can be turned off in settings. However, it is not recommended to enter sensitive data such as business contracts, personal ID numbers, and bank card numbers into any AI service. Local deployment is the only insurance for zero data transmission.

Why DeepSeek is so much cheaper than ChatGPT

Three reasons. The first is that the labor cost and electricity cost of the Chinese team are 50% lower than those in Silicon Valley. The second is that MoE hybrid expert architecture inference only activates some parameters and has low computational overhead. The third is a large-scale revenue strategy. DeepSeek does not pursue short-term profits but captures market share. Prices are expected to increase by 30% to 50% in 2027 but will still be significantly lower than OpenAI.