DeepSeek in French: free, secure, or local ?

#DeepSeek in French: the short answer

DeepSeek refers to two things that must be separated from the outset. On one side, an online service: a chat website and mobile app operated by the Chinese company DeepSeek on its servers. On the other, a family of models whose weights are published on Hugging Face and that anyone can download and run locally. The name is the same, but the implications for your data are opposite.

Free?

Yes for discussion on the official website and in the official app, with an account. API access, intended for developers, is usage-based and outside the scope of this guide.

In French?

Yes. The model understands and writes French: just write your question in French. There is no special “DeepSeek French” version to find or install.

Sure?

That depends on what you write there. The official privacy policy states that collected data, including your messages and files, is stored on servers located in the People’s Republic of China and may be used to improve the service.

Locally?

Possible and free with Ollama. However, the versions that run on a typical PC are smaller models than the online service's model: you gain privacy but lose some quality.

The rest of this guide covers each point in detail. If you only want to ask low-stakes questions, the next section is enough. If you plan to paste in a contract, client file, or payslip, read the privacy section first.

#Where to use DeepSeek for free

The official entry point is DeepSeek's chat website, accessible from a browser, and the same publisher's mobile app on the App Store and Google Play. Both require you to create an account. The chat itself requires neither a credit card nor a subscription.

https://chat.deepseek.com

1. 01
   Open the official website

   Type the address yourself instead of clicking an ad. DeepSeek's success has led to many websites and applications using its name without any connection to the publisher. Official domains end in deepseek.com.
2. 02
   Create an account

   Sign-up uses an email address or a third-party account. You do not have to use your work address: a dedicated address limits what the service knows about you.
3. 03
   Write in French

   Ask your question directly in French. The model responds in the language of the message. If a response switches to English, simply add « Respond in French » to your request.
4. 04
   Choose the response mode

   The interface offers a deep-reasoning mode (the DeepThink button), which is slower but more rigorous on multi-step problems, as well as web search. For rewriting or summarizing, normal mode is sufficient.

Free doesn't mean unlimited. The service is regularly overloaded during peak hours and may then reject a request, indicating that the server is busy. DeepSeek does not publish, to our knowledge, a stated quota for free chat, so we don't provide one. If you need guaranteed availability, look at the paid API or local execution, not the free interface.

#French: interface and response quality

You need to distinguish between the interface language and the model's language. The interface—that is, the menus and buttons—follows the languages the publisher has chosen to translate and may appear in English depending on your device. The model, however, was trained on text in many languages, including French. That's what matters: an English interface doesn't prevent you from getting answers in French.

We do not publish a “French quality” rating for the online service: we have no reproducible measurement to present, and a figure without a protocol would be meaningless. What we can say without making anything up is what limitations to watch for—limitations shared by large models trained primarily on English and Chinese.

Local references

French law, taxation, administrative procedures, school curricula: the model may answer confidently based on rules from another country or outdated rules. Always verify against an official French source.

Displayed reasoning

In reasoning mode, intermediate steps may appear in a different language from the final answer. This says nothing about the quality of the conclusion, but it can be surprising.

Register and typography

Mixed informal and formal address, English quotation marks, phrasing modeled on English. An explicit instruction at the start of the conversation (“formal address, French from France, professional tone”) fixes most of it.

Sensitive topics

The online service applies moderation rules set by its publisher. Some political questions may remain unanswered. This is a service behavior to be aware of before using it for monitoring or research.

#What happens to your text on the online service

The reference here is the privacy policy published by DeepSeek (link at the end of the guide). The points below summarize the version dated February 14, 2025. This text changes over time: the date shown at the top of the official page is authoritative, and if it is more recent than this one, reread the relevant sections before deciding.

What is collected

Account information (email address, phone number if applicable), the content you enter (messages, uploaded files, conversation history, feedback), and technical information about the device and connection, including the IP address.

What it’s for

To provide the service, keep it secure, and improve and train the publisher’s technology. Your conversations are therefore not processed only for the duration of the response.

Where it is stored

The policy states that the collected information is stored on servers located in the People’s Republic of China.

Who it’s shared with

Technical service providers, group companies, and authorities when the law applicable to the publisher requires it.

Your levers

Delete conversations or the account from the settings, and exercise your rights of access and deletion with the publisher, whose contact information appears in the policy. Also check whether your account settings let you disable the use of your conversations to improve the model.

None of this is hidden: it is stated by the publisher itself. The practical consequence is simple. Everything you paste into the chat window leaves your device, goes to servers outside the European Union, and may remain there. For a recipe or a math exercise, this does not matter at all. For data that does not belong to you, it is a different matter.

If you use DeepSeek in a professional setting, this is not an end-user issue. Sending personal data belonging to customers, employees, or patients to a service located outside the European Union is a transfer under the GDPR, which must be governed and documented by the data controller. This is a decision for your data protection officer or management, not an application setting. Until they provide guidance, the prudent rule fits on one line: nothing identifying, nothing confidential, nothing covered by professional secrecy.

This issue is not specific to DeepSeek. Every online assistant, including ChatGPT, receives what you write and applies its own retention and reuse policy. What differs from one service to another is the hosting country, the opt-out options offered, and the contractual commitments available to businesses. Comparing DeepSeek and ChatGPT on this basis therefore means reading two privacy policies, not two score tables.

#Online or local: how to decide

The choice is not about the technology but about the nature of what you're going to write. Ask yourself before each use: if this text were published tomorrow, would it pose a problem for you or someone else?

Curiosity, general knowledge, homework, non-sensitive code

The free online service is sufficient. You get the provider's most capable model without installing anything.

Personal text you’d rather keep private

Mail, correspondence, health, household finances: favor a local model. Quality will be slightly lower, but nothing leaves your machine.

Customer, employee, student, or patient data

No public-facing online service without internal validation. A local model or hosting controlled by your organization is the normal approach.

Documents covered by secrecy or confidentiality clauses

Local only, on a machine you administer. This use case alone justifies buying a graphics card.

Need the best possible quality on non-sensitive data

Online. No desktop PC can run the largest DeepSeek model under good conditions.

Many readers end up using both: the online service for low-stakes questions and a local model for anything involving their documents. This is not a contradiction; it is the most sensible division of labor.

#DeepSeek locally: the essentials before installation

DeepSeek publishes its model weights on its Hugging Face page. Once a model is downloaded, it runs on your processor or graphics card: your questions don't pass through any server, and the online service's privacy policy no longer applies because you are no longer using that service. You can even disconnect from the network while using it.

The trade-off is size. The models powering the online service have several hundred billion parameters and require workstation- or server-class hardware. What you install on an ordinary PC are the so-called distilled versions of DeepSeek R1: models ranging from 1.5 to 70 billion parameters, built on Qwen or Llama base models and trained to imitate the larger model's reasoning. They carry the name DeepSeek, reason visibly, but do not give the same answers as the website.

deepseek-r1:7b

About 5 GB of VRAM in Q4_K_M. Runs on a RTX 3060 12 GB or a recent Mac. Useful for exploring, but limited in formal French.

deepseek-r1:14b

About 9 GB of VRAM in Q4_K_M. The right starting point for a 12 GB card (RTX 3060, RTX 4070).

deepseek-r1:32b

Approximately 19 GB of VRAM in Q4_K_M. Requires a RTX 4090 24 GB or a Mac with a comfortable amount of unified memory (M4 Pro 24-48 GB).

deepseek-r1:70b

About 40 GB of VRAM in Q4_K_M. Beyond the reach of a single consumer graphics card.

These memory figures are the sizing benchmarks used on this site for Q4_K_M quantization, not speed measurements. Actual throughput depends on your hardware and context length; we do not state it here.

1. 01
   Install Ollama

   Ollama is the program that downloads and runs models. It is available for Windows, macOS, and Linux and installs from ollama.com. Once started, it listens on http://localhost:11434.
2. 02
   Choose the size based on your graphics memory

   Go back to the list above. When in doubt, choose the smaller size: a model that fits entirely in VRAM is much more pleasant to use than a larger model that spills into RAM.
3. 03
   Run the model

   A single command downloads the model on first launch, then opens a conversation in the terminal. Write in your language, just like on the site.
4. 04
   Add an interface if needed

   To get a chat window similar to the online service, Open WebUI or LM Studio connect to Ollama. Everything remains on your machine.

# Télécharge le modèle au premier lancement, puis ouvre la discussion
ollama run deepseek-r1:14b

# Machine plus modeste : la version 7B
ollama run deepseek-r1:7b

To verify that everything is working properly locally, query your machine’s Ollama service directly. The localhost address never leaves your computer.

curl http://localhost:11434/api/generate -d '{
  "model": "deepseek-r1:14b",
  "prompt": "Réponds en français : explique en trois phrases ce que fait la quantification.",
  "stream": false
}'

We intentionally keep the installation section brief: it is covered step by step, along with context and temperature settings, in the dedicated guides listed at the end of the page. The family's major recent models, such as DeepSeek V4 Flash, also have their own guide, including the hardware they require.

#Pitfalls and misconceptions

“Locally, I get the same DeepSeek as on the website”

No. On a typical PC, you run a smaller, distilled version. Expect less precise answers, especially for long texts and polished French.

“The local model still sends my data to China”

A weight file is not a program that communicates: it is a set of numbers read by Ollama on your machine. To verify this, disconnect from the network after downloading and confirm that the model still responds.

“A third-party application DeepSeek, it’s the same thing”

No. It adds an intermediary whose hosting and policies you don't know. Stick to the official website, the publisher's app, or the local option.

“The displayed reasoning proves that it’s correct”

The reasoning text shows a line of thought, not verification. A well-presented chain of reasoning can lead to a false conclusion. Check the facts, figures, and references.

“Deleting the conversation erases everything”

Deleting from the interface removes the conversation from your history. What happens to server-side copies depends on the retention periods described by the provider, which you should read in its policy.

“The local model censors just as much as the website”

Some of the site's moderation is enforced by the service around the model. Locally, this layer disappears, but the model retains the biases acquired during training. The behavior differs, without becoming neutral as a result.

#Official sources to consult

The claims in this guide about the online service refer to the vendor's documentation. Consult it directly: it is updated without notice, and its revision date appears at the top of the page.

https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html

https://api-docs.deepseek.com/

https://huggingface.co/deepseek-ai

For the local side, each model page on Hugging Face lists its license and size: that’s where you need to check the terms of use before professional deployment, model by model.

#Go further

This guide helps you choose between an online service and local deployment. To put it into practice, these site guides pick up where it leaves off:

Install DeepSeek R1 with Ollama

Step-by-step installation of the 7B, 14B, and 32B distilled versions for beginners. https://quelllm.fr/guide/installer-deepseek-r1-ollama

Configure DeepSeek with Ollama

Temperature, context, and memory settings once the model is installed. https://quelllm.fr/guide/configurer-deepseek-ollama-guide

DeepSeek V4 Flash

The family’s major recent model and the hardware it actually requires. https://quelllm.fr/guide/guide-deepseek-v4-flash

Local LLM and GDPR

The framework to know before processing personal data with AI, online or locally. https://quelllm.fr/guide/llm-local-donnees-privees-rgpd