What is DeepSeek V4 and why is it different from previous models

The Chinese startup DeepSeek made its mark on the world AI map less than two years ago as a team capable of doing exceptional things with significantly fewer resources than the tech giants from Silicon Valley. DeepSeek V4 is the logical culmination of this philosophy — a model that on paper looks like a colossal machine (a total of 1.6 trillion parameters), but in reality, activates only a fraction of them with each response.

Behind this is the MoE (Mixture of Experts) architecture — in Czech, "směs expertů". Instead of engaging all parameters at once, the model selects only the part of the neural network most relevant to the given task for each query. In practice, this means that DeepSeek V4-Pro activates only 49 billion parameters during inference from the total trillion-parameter volume. The result is the performance of a large model at the cost of a small one — and that is precisely the trump card DeepSeek is playing.

The model also brings technical innovations: a new hybrid attention architecture combining CSA (Compressed Sparse Attention) and HCA (Heavily Compressed Attention), which makes working with long contexts 73% more efficient compared to its predecessor DeepSeek-V3.2. In addition, it introduces the Muon optimizer for more stable training and the mHC (Manifold-Constrained Hyper-Connections) mechanism for better signal transfer across model layers. The technical report was published on arXiv.org.

Two variants: Flash for speed, Pro for performance

DeepSeek V4 is not one model, but a family:

• DeepSeek-V4-Flash — 158 billion parameters, designed for fast responses and everyday tasks. Activating only a part of the network makes it one of the most economical choices on the market.
• DeepSeek-V4-Pro — 862 billion active parameters (out of a total of 1.6 trillion), designed for complex analyses, intricate coding, and scientific tasks.

Both models support a context window of 1 million tokens — which corresponds to approximately 750,000 words or an entire book. The maximum output length reaches 384,000 tokens. For comparison: GPT-4o handles 128,000 tokens, Claude 3.5 Sonnet 200,000 tokens.

Three thinking modes: from speed to maximum

One of the most interesting innovations of DeepSeek V4 is the system of three reasoning modes:

• Non-Think — fast, intuitive answers without deep analysis. Suitable for routine questions and quick searches.
• Think High — engages logical analysis and conscious reasoning. Great for complex problems.
• Think Max — maximum performance. The model thinks longer, but the results are significantly better for difficult tasks.

The difference between the modes is measurable. On the LiveCodeBench benchmark (tests programming on real-world tasks), V4-Pro achieves 56.8 points in Non-Think mode, while in Think Max, it jumps to 93.5 points. On the IMOAnswerBench mathematical olympiad, the jump is even more dramatic: from 35.3 to 89.8 points.

Benchmarks: where DeepSeek V4 leads and where it lags

DeepSeek V4-Pro (in maximum Think Max mode) was compared with the best models on the market — Claude Opus 4.6, GPT-5.4, and Google Gemini 3.1 Pro. The results are surprising:

Programming — DeepSeek dominates

On LiveCodeBench, DeepSeek V4-Pro Max achieves a score of 93.5, while Gemini-3.1-Pro stands at 91.7. In competitive coding on Codeforces, it achieves a rating of 3206 — a value corresponding to international programming masters. On the SWE Verified benchmark, which tests real software bug fixes, it achieves 80.6%, and on BrowseComp (autonomous web browsing) 83.4%.

Knowledge and Reasoning — competition sometimes leads

On the GPQA Diamond benchmark (doctoral questions in physics, chemistry, biology), DeepSeek achieves 90.1 points — a decent number, but Gemini-3.1-Pro leads with 94.3 and GPT-5.4 achieves 93.0. On SimpleQA-Verified, the situation is similar: DeepSeek 57.9 vs. Gemini 75.6.

In other words: in programming and agent tasks, DeepSeek V4 is at the top or very close to it. In factual knowledge and scientific reasoning, Gemini and GPT still sometimes surpass it.

Price: a fraction of what you pay elsewhere

And here comes what makes DeepSeek DeepSeek. API access prices are aggressively low:

• V4-Flash: input $0.14 / million tokens, output $0.28 / million tokens
• V4-Pro: input $0.435 / million tokens, output $0.87 / million tokens

For comparison: Claude Opus 4.6 costs $15 per million input tokens and $75 per million output tokens. GPT-4o is around $2.50 for input. DeepSeek V4-Pro is approximately 3–17× cheaper than comparable Western models in typical use.

The web version at chat.deepseek.com is freely available from the Czech Republic without the need for a VPN.

Open Source Code: an advantage for developers in the Czech Republic and Europe

DeepSeek V4 is open source under the MIT license — which is crucial for the developer community. The model weights are freely available on Hugging Face, where over 1.2 million users have downloaded them as of the publication date of this article.

Thanks to this, companies and developers — including Czech ones — can run DeepSeek V4 on their own hardware without sending data to external servers. This is a significant advantage for GDPR compliance. The smaller V4-Flash variant (158B parameters) is also available for organizations with limited hardware. Deployment is possible via frameworks like vLLM or SGLang.

Czech startups and research institutions looking for a powerful model without high API costs and without dependence on American clouds now have a very concrete alternative.

Availability in Czech

DeepSeek V4 handles Czech. The model was trained on multilingual data and works well for factual queries, translations, and programming. Creative writing and idiomatic expressions may be less natural than with models primarily trained on European data, but for technical and analytical tasks, Czech is fully functional.

The web version does not have a Czech interface localization — you communicate with the model in Czech, but the menus and settings are in English and Chinese.

Security and GDPR: what to consider

When using the public web version, the standard warning applies: data may be used for model training. DeepSeek's server infrastructure is Chinese, which is a sensitive topic in the EU regulatory environment — especially after the introduction of the EU AI Act.

For corporate deployment in the Czech Republic and the EU, the recommended approach is either an API with clearly defined data processing conditions, or — and this is true freedom — local deployment of open source weights on your own server.