DeepSeek: Chinese AI Provider with V3, R1 and V4 Models Overview

Hardly any AI company has shaken up the industry over the past year and a half as much as DeepSeek. When the Chinese company introduced its DeepSeek-R1 model in early 2025, which could compete with the most expensive American AI models on many tasks while allegedly costing only a fraction to train, technology stock markets worldwide wobbled. Since then, a lot has happened: DeepSeek has released new model generations, its chat app is among the most downloaded AI applications worldwide, and at the same time the provider is under intensive review by German and European data protection authorities. For case workers, self-employed people and private individuals wondering whether DeepSeek is a sensible addition to their everyday work, it’s worth taking a factual look at its origin, offerings, models, pricing and, above all, the data protection question – as part of our overview of AI models compared.

Who is behind DeepSeek?

• DeepSeek website
• Open DeepSeek Chat directly

DeepSeek was founded in Hangzhou in 2023 and is closely linked to the Chinese hedge fund High-Flyer Capital Management, out of whose data center infrastructure the company originally emerged. Founder Liang Wenfeng, a graduate of Zhejiang University, had already built a company with High-Flyer before founding DeepSeek that developed quantitative trading strategies using machine learning. The experience gained there in the efficient use of computing power fed directly into the later AI development.

DeepSeek’s stated goal is the development of what is known as Artificial General Intelligence – an AI that matches or surpasses human thinking in as many areas as possible, rather than merely solving individual tasks well as previous systems did. Unlike many Western providers, DeepSeek consistently pursues an open-weight strategy: the model weights are published and made available for download and self-use. This openness, combined with strikingly low training and usage costs, was the actual trigger for the worldwide stir in January 2025.

DeepSeek showed the industry that powerful AI models don’t necessarily require billion-dollar budgets – that made the provider famous overnight, but it also raises questions about data origin and security.

Lukinski AI · ab 49 €/MonatFrag die Immobilien-KI — Jetzt testen →Frag z.B.: „Ist diese Mietvertragsklausel wirksam?“

DeepSeek’s products at a glance

The chat app

For private users getting started, DeepSeek offers a free chat application, available as a website as well as an app for iOS and Android. It works like familiar chatbots: you type in a question or task and receive a text answer, optionally with a visible “thinking process” if a reasoning model is activated. The app also supports file and image uploads for analysis as well as a web search feature that can incorporate current information.

The API for businesses and developers

For companies, agencies and developers, DeepSeek provides a programming interface (API) through which custom applications, chatbots or automated text processing can be connected to the DeepSeek models. API usage is billed according to the amount of text processed, measured in so-called tokens – roughly comparable to word fragments. This is exactly where DeepSeek’s main advantage over many competitors lies.

The DeepSeek model family at a glance

Anyone who looks more closely at DeepSeek quickly comes across a whole series of model names: V3, V3.1, V3.2, R1, V4 Flash, V4 Pro. To outsiders this quickly looks confusing, yet there is a clear idea behind each version. DeepSeek releases its models at relatively short intervals and usually numbers new versions sequentially. Broadly, the models can be divided into two lines: the classic V series (V3, V3.1, V3.2, now V4) for general text tasks, and the R series (currently R1) for tasks where the AI is meant to derive its answer through traceable intermediate steps, a so-called reasoning model.

Model	Type	Key feature	Typical use
DeepSeek V3	Base model	Introduction of the current efficiency architecture	General text tasks
DeepSeek V3.1 / V3.2	Base model (further developed)	Improved accuracy, better tool use (“agent” capabilities)	Research, text processing, simple automation
DeepSeek R1	Reasoning model	Visible, multi-step thinking process before the answer	Mathematics, logic, complex analyses
DeepSeek V4 Flash	Base model (fast)	Very low cost, very large text volumes possible	Bulk processing, simple everyday tasks
DeepSeek V4 Pro	Base model (flagship)	Highest accuracy of the current generation	Demanding professional and business tasks

What does “open weight” mean, and why does it matter?

The Mixture-of-Experts architecture explained simply

Most DeepSeek models use a technique called Mixture-of-Experts. You can picture it like a large team of specialists: instead of consulting the entire team for every request, the system automatically selects only the few experts actually needed for the specific task. As a result, a single request consumes considerably less computing power than with models that always activate the entire system – this is the main reason for the low prices.

What the MIT license means for users

DeepSeek publishes the so-called model weights (the model’s “trained settings”) openly and under the MIT license. This means: anyone – including companies – may download the models, run them on their own servers, adapt them and even reuse them commercially, without paying license fees. For technically capable businesses, this opens up an interesting possibility to use the models’ capabilities without sending data to DeepSeek servers in China, if the model is instead run on their own or on European infrastructure.

Rule of thumb: DeepSeek’s underlying model technology can fundamentally be used in a more privacy-friendly way than the standard chat service – provided you run the model yourself and don’t send data to China.

DeepSeek R1: the reasoning model

DeepSeek-R1 was the model that made DeepSeek famous worldwide in early 2025. The difference from classic language models: R1 visibly “thinks” through several steps before giving a final answer – similar to a person solving a math problem step by step on a scratch pad instead of directly guessing the result. You can even have these intermediate steps displayed in the chat app. As a result, R1 achieves considerably better results than purely “instant answer” models on tasks with clear logic – such as mathematics, debugging code, or multi-step text analyses. The downside: thinking through several steps takes more computing time and is therefore more expensive for the same amount of text than the simpler V models.

DeepSeek V4: the current flagship

In spring 2026, DeepSeek followed up with the V4 generation, offered in two variants.

Variant	Input price (per 1 million tokens)	Output price (per 1 million tokens)	Context window
V4 Flash	approx. 0.14 US dollars	approx. 0.28 US dollars	up to 1 million tokens
V4 Pro	approx. 0.44 US dollars	approx. 0.87 US dollars	large, designed for complex documents

The “context window” describes how much text the model can keep “in view” at once – a large context window makes it possible, for example, to process an entire book or several long contract documents in a single pass.

Pricing models: why DeepSeek is so inexpensive

Since its market launch, DeepSeek has deliberately positioned itself as the most affordable alternative among the major AI providers. This is made possible by a particularly efficient model architecture (a so-called Mixture-of-Experts structure, in which only a small part of the overall model actually computes for each request) as well as by an aggressive pricing policy in the competition for developers.

Model	Input price (per 1 million tokens)	Output price (per 1 million tokens)	Area of use
DeepSeek V4 Flash	approx. 0.14 US dollars	approx. 0.28 US dollars	Fast everyday tasks, high text volume
DeepSeek V4 Pro	approx. 0.44 US dollars	approx. 0.87 US dollars	More complex tasks, higher accuracy
DeepSeek R1 (reasoning)	approx. 0.70 US dollars	approx. 2.50 US dollars	Tasks with step-by-step logic

For comparison: comparable models from established US providers in a similar performance class are often priced many times higher. New customers at DeepSeek also receive a quota of free tokens, and repeated, similar requests are billed considerably cheaper via caching. For private users, the chat app is free to use anyway; only very high usage volumes or extended features incur costs.

Calculation example: the cost of a typical task

A case worker has a 30-page document (about 15,000 words, roughly 20,000 tokens) summarized and receives a summary of about 1,000 words (around 1,300 tokens) back.

• With V4 Flash: 20,000 input tokens x 0.14 US dollars / 1 million = 0.0028 US dollars, plus 1,300 output tokens x 0.28 US dollars / 1 million = 0.00036 US dollars. Total cost: under 0.01 US dollars.
• With V4 Pro: 20,000 input tokens x 0.44 US dollars / 1 million = 0.0088 US dollars, plus 1,300 output tokens x 0.87 US dollars / 1 million = 0.00113 US dollars. Total cost: around 0.01 US dollars.

Even with a hundred such summaries a day, the pure API costs are therefore in the range of a few euros a month – the real effort usually lies in the technical integration, not in the model price.

Data protection and GDPR: the central question

Anyone using DeepSeek via the official app or the standard API sends their inputs to servers operated in China. This is the decisive difference from many other AI providers and the point German users should absolutely understand before using it.

Why the server location China is legally relevant

The General Data Protection Regulation (GDPR) allows the transfer of personal data to countries outside the EU only under certain conditions. For China, there is no so-called adequacy decision from the EU Commission that would certify a level of data protection comparable to that of the EU. In addition, Chinese laws – in particular the Cybersecurity Law and the Data Security Law – require companies based in China to grant Chinese authorities access to data upon request, even without the affected person being informed or a German court being able to order it. It is precisely this state right of access that makes it legally very difficult to conclude a so-called transfer impact assessment positively – a review with which a company must demonstrate, before transferring data to a country outside the EU, that the data is sufficiently protected there, as required by the GDPR for data transfers to unsafe third countries.

What this means for you in practice

For private individuals using DeepSeek for uncritical, everyday questions (such as wording help, general research or translations without sensitive content), the legal risk is manageable, since this is usually not processing in a business sense. The picture changes as soon as personal data belonging to third parties, trade secrets, customer data, health data or financial information is entered. Companies and self-employed people who want to use DeepSeek for business purposes actually need a data processing agreement (DPA) as well as a named representative of the provider in the EU – as things currently stand, DeepSeek does not adequately provide either. Several German state data protection authorities have been conducting formal review proceedings on this since 2025, and use of the official chat app continues to be viewed critically or restricted in parts of Germany.

A more privacy-friendly alternative

Important to know: since DeepSeek publishes its models openly as so-called open-weight models, technically capable companies can also run the models on their own hardware or with a European cloud provider. In this case, the data never leaves the company’s own or European infrastructure in the first place, and the third-country problem is eliminated entirely. For most case workers and smaller businesses, however, such self-hosting is hardly realistic technically or economically.

Rule of thumb: for DeepSeek in its standard form – private, uncritical questions are unproblematic; personal or business-sensitive data has no place in the input field.

Strengths and weaknesses at a glance

Strengths	Weaknesses
Very low prices for API use	Server location China, unclear GDPR situation
Strong performance on logic and programming tasks	No data processing agreement for businesses
Models open and self-hostable (open weight)	No German-language user interface in every area
Free chat app with a broad feature set	No native image or video generation tool

Who is DeepSeek suited for?

DeepSeek suits cost-conscious users without sensitive data.

• Private individuals drafting texts
• General research and learning aids
• Programming support in a private context
• Developers focused on price

For real estate agents, property managers and companies with customer data, leases or listing exposés, the standard offerings are best avoided. For legally sound processing with German hosting, Lukinski AI is better suited.

Practical examples from everyday life and work

1. Drafting and revising texts

V4 Flash or V3.2 are well suited to drafting and polishing the language of emails, cover letters or social media posts – unproblematic for private purposes.

2. Programming and debugging

R1 and V4 Pro are often used to help with program code, for example to explain errors or complete code sections.

3. Research and summaries

Long trade articles, studies or news situations can be summarized in a single pass thanks to the large context window, which can save case workers a lot of reading time.

4. Analyzing real estate and financial documents (with reservations)

Technically, a DeepSeek model could also analyze listing exposés, leases or financing documents and extract key points. However, since such documents almost always contain personal data belonging to tenants, buyers or borrowers, use via the standard DeepSeek API is strongly discouraged unless a GDPR-compliant contractual basis and a European processing location are ensured. Anyone looking for a legally sound AI solution with German hosting for such use cases will find a more suitable route with Lukinski AI.

5. Customer service chatbot for uncritical inquiries

For simple, non-personal standard questions (such as opening hours or general product information), a DeepSeek model can be integrated cost-effectively as a chatbot foundation, provided no customer data is processed in the conversations.

Calculation example: how inexpensive DeepSeek really is in everyday use

A case worker has about 200 short customer inquiries drafted per month with DeepSeek V4 Flash, each with an estimated 500 input tokens and 700 output tokens.

• Input: 200 x 500 = 100,000 tokens, at 0.14 US dollars per million tokens that comes to around 0.014 US dollars
• Output: 200 x 700 = 140,000 tokens, at 0.28 US dollars per million tokens that comes to around 0.039 US dollars
• Total cost per month: under 0.10 US dollars

This calculation example shows why DeepSeek is considered especially inexpensive for uncritical text volumes – the actual decision should nevertheless be based on data protection, not on price.

Rule of thumb: price and performance speak in favor of DeepSeek – but for business use with sensitive data, data protection should always be the primary basis for the decision, not cost savings.

Limitations and open points

Despite the impressive price-performance ratio, there are limitations: DeepSeek so far offers no mature image or video generation of its own, unlike some competitors; integration with third-party software (plugins, office integrations) is less developed; and the data protection situation when using the standard services remains unresolved and continues to be reviewed by German authorities.

Frequently asked questions about DeepSeek

Is DeepSeek banned in Germany?

There is no nationwide, general ban. However, several German state data protection authorities are reviewing its use critically, and in individual areas – such as on government devices or in certain federal states – restrictions or recommendations against use already exist.

Can I use DeepSeek for free?

Yes, the chat app is free for private use. For API use via your own applications, fees apply according to the amount of text, though new customers receive a free starting quota.

Does DeepSeek store my inputs?

Yes, as with most chat AIs, inputs and chat histories are stored and, according to the privacy notices, used among other things to improve the models. With DeepSeek, the storage location is on servers in China.

Is DeepSeek better than ChatGPT?

That can’t be answered across the board. In independent tests, the models are close to each other on many tasks; DeepSeek scores especially on price, while other providers are sometimes ahead on data protection, additional features, or image generation.

What is the difference between V4 Flash and V4 Pro?

Flash is cheaper and faster, but somewhat less precise on complex tasks. Pro delivers higher accuracy, but costs about three times as much per unit of text processed.

Can I run a DeepSeek model myself?

Yes, since the models are published openly under the MIT license, they can be run on sufficiently powerful own or rented server hardware – however, this requires technical expertise and a corresponding investment in computing power.

Is R1 the same as ChatGPT with “thinking mode”?

Similar in basic principle – both show a multi-step path to the solution before the actual answer – but the specific technical implementation and the underlying training data differ between providers.

Other AI providers compared are summarized on the comparison page of all AI providers.