A Million Tokens as Standard. This Isn't the Future, It's Now

The most striking new feature of DeepSeek-V4 is support for a context length of one million tokens (1M) in both model variants. For perspective: one million tokens corresponds to roughly 750,000 words, or several novels, or the complete source code of a large software project. Until now, such capacities have been the privilege of a handful of the most expensive closed models, often only in experimental mode.

DeepSeek promises that the 1M context will be a standard across all official services — the web interface, mobile app, and API. Technically, this is enabled by a new Hybrid Attention architecture, which combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). Thanks to this, V4-Pro requires only 27% of the computational power and 10% of the KV cache memory compared to the previous V3.2 generation when processing a million-token context. In layman's terms: the model can "read" massive documents faster and cheaper, without forgetting what was at the beginning.

Long-context benchmarks confirm this. In the MRCR 1M test, V4-Pro-Max achieved a score of 83.5%, while Gemini-3.1-Pro High remained at 76.3%. This means the model remembers details scattered throughout long texts better — crucial, for example, for legal analysis, code audits, or scientific paper research.

Two Models for Two Worlds: Pro for Demanding Tasks, Flash for Everyday Work

DeepSeek-V4 comes in two variants that differ in size, price, and focus:

DeepSeek-V4-Pro is the flagship. It has 1.6 trillion parameters, with 49 billion activated per query. In Think Max mode (maximum reasoning effort), it achieves results comparable to the best closed models in the world. In the programming benchmark LiveCodeBench, it scored 93.5%, surpassing Gemini-3.1-Pro (91.7%) and Claude Opus 4.6 Max (88.8%). On the competitive platform Codeforces, it achieved a rating of 3206, which is an elite level. In the mathematical test IMO Answer Bench, it solved 89.8% of problems, just behind GPT-5.4 xHigh (91.4%).

DeepSeek-V4-Flash is more compact: 284 billion parameters, 13 billion activated. It is faster and cheaper, yet retains the ability to solve complex tasks, especially when Think Max mode is enabled. In simple agentic tasks, it performs almost as well as the Pro version. For Czech startups and developers, it can be an ideal choice for integration into their own applications.

Agents, Code, and Tools: The Main Battleground of 2026

DeepSeek-V4 doesn't fall into the "chatbot only" category. The model is specifically optimized for agentic tasks — that is, the ability to independently complete complex assignments using tools. In the Terminal Bench 2.0 test, which simulates command-line work, V4-Pro-Max achieved 67.9%. In SWE Verified, a benchmark for solving real software bugs, it handled 80.6% of cases, which is on par with Claude Opus 4.6 Max (80.8%).

The model is directly adapted for popular agentic frameworks such as Claude Code, OpenClaw, OpenCode, and CodeBuddy. It supports tool calls, structured JSON output, and even offers compatibility with the Anthropic API format — which simplifies the transition for developers who have been using Claude.

In practice, this means V4 can not only write code, but also navigate repositories, generate documentation, create presentations, or browse websites to collect data. In internal DeepSeek tests, V4-Pro is reportedly used as the main model for agentic programming, and the results allegedly surpass Claude Sonnet 4.5.

Prices That Change the Game

DeepSeek has traditionally pushed API prices down, and V4 is no exception. For one million input tokens, you'll pay just 0.14 USD for V4-Flash (0.028 USD on cache hit), and 0.28 USD for output tokens. V4-Pro costs 1.74 USD per million input tokens (0.145 USD on cache hit) and 3.48 USD for output. For comparison: OpenAI GPT-5.4 or Claude Opus 4.6 often cost several times more, while their context windows are shorter.

The web interface at chat.deepseek.com and the official mobile app are free for end users. The API is available at platform.deepseek.com. Developers should be aware of a planned change: the current model names deepseek-chat and deepseek-reasoner will be discontinued on July 24, 2026, and replaced directly by V4-Flash.

What Does This Mean for the Czech Republic and Europe?

For Czech users and companies, DeepSeek-V4 brings several key advantages and questions. Because it is an open-source model with an MIT license, Czech companies can run it locally on their own servers. This eliminates the risk of sensitive data leaking abroad and facilitates compliance with GDPR. The models are available on Hugging Face and Chinese ModelScope.

Regarding language support, DeepSeek, as a Chinese model, has historically also mastered Czech at a very decent level — previous V3 and R1 generations handled Czech texts, translations, and grammar well. However, the official documentation does not explicitly mention Czech, so the most demanding tasks may still be better handled by models from OpenAI or Google.

At the European level, the EU AI Act must be taken into account. The Chinese origin of the model may raise questions regarding transparency of training data, copyright, and potential censorship. While open weights and the technical report provide a high degree of transparency, companies in regulated industries should carry out their own risk assessment according to the AI Act classification.

Comparison with the Best: Where V4 Wins and Where It Falls Short

In benchmark tables, DeepSeek-V4-Pro-Max often stands alongside the best closed models. In coding and mathematics, it is at the top; in some agentic scenarios, it even surpasses Claude Opus 4.6. In general knowledge, measured for example by the MMLU-Pro test, it remains slightly behind Gemini-3.1-Pro (87.5% vs. 91.0%). In the Apex test, which measures deep scientific understanding, Pro-Max scores 38.3%, while Gemini-3.1-Pro reaches 60.9%.

But this doesn't mean V4 is weak. Rather, it shows that every model has different strengths. For developers, data analysts, and companies looking for an economical alternative to Western APIs, DeepSeek-V4 can be a significantly more attractive choice than more expensive competitors.

Conclusion: An Open-Source Alternative That Is No Longer "Just an Alternative"

DeepSeek-V4 pushes the boundaries of what can be expected from an open-source model. A million-token context, top-tier coding results, competitive prices, and open weights — all of this makes V4 a serious player that can save Czech companies tens to hundreds of thousands of crowns annually on API costs. At the same time, it raises new questions about security, data, and the geopolitical divide of the AI ecosystem. One thing is certain: 2026 will belong to models that can not only answer, but also act independently.