The new generation of language models from DeepSeek, the Chinese startup based in Hangzhou, has become one of the most talked-about developments in the artificial intelligence landscape. Their series DeepSeek V4 bursts in At a time of saturation of offers and promises, but with a very clear message: to offer top-level capabilities in reasoning and agents, with a gigantic context window and, above all, with a much lower cost of use than its American rivals.
This launch is not an isolated event. It comes just a few days after new versions of ChatGPT and other closed modelsAnd amid China's ongoing efforts to reduce its technological dependence on the West, DeepSeek V4 combines a open source strategyAggressive pricing and a close alliance with Huawei in the hardware field could shift the balance of power for European companies seeking cheaper and more controllable alternatives.
What is DeepSeek V4 and how is it built?
The DeepSeek V4 family consists of Mixture-of-Experts (MoE) language models designed to be highly efficient in long-term contexts. The company speaks of a main model with around 1 trillion total parameterswhere only a fraction is activated in each token, and an even more ambitious version, V4-Pro, which reaches the 1,6 billion parameters, with approximately 49.000 billion in assets per query.
The key to DeepSeek's MoE approach is that, although the total number of parameters is enormous, only tens of billions are activated at each inference step. This allows drastically reduce the cost of computing and memory compared to dense models of similar size. Some technical materials mention approximately 37.000 billion active parameters per token In one of the variants, this fits with the idea of extreme efficiency without sacrificing performance.
Two variants: V4-Pro and V4-Flash for different needs
DeepSeek has segmented its offering into two main models: V4-Pro and V4-FlashThe V4-Pro is the flagship model, with the aforementioned 1,6 trillion parameters and a focus on deep reasoning, global knowledge, and agentic capabilitiesespecially in programming and complex analysis tasks.
For its part, V4-Flash is presented as the light and fast versionIt has approximately 284.000 billion total parameters and around 13.000 billion active ones, also maintaining a context window of up to one million tokensIt is designed for massive deployments, where the cost per token, latency, and scalability are more important than squeezing every last tenth out of benchmarks.
This Pro/Flash duality has more than commercial strategy that goes beyond simple technical differentiationV4-Pro targets large corporations, laboratories, and use cases where the highest possible performance is required; V4-Flash, on the other hand, aims to be the "workhorse" for SMEs, startups, and high-volume services that need competent and inexpensive AI.
Context of a million tokens: what changes in real-world use
One of the most striking features of DeepSeek V4 is its ultra-long context window of up to one million tokensThis is well above the 128.000 tokens typical in many current business models. In practice, this equates to being able to work with hundreds of thousands of words in a single query: complete technical manuals, large code repositories, legal or historical customer service records without needing to fragment them.
Beyond the figure, this broad context can reduce costs and errors resulting from splitting documentsFewer API calls are needed, there's less risk of losing information between scans, and the design of products that rely on large volumes of text is greatly simplified. DeepSeek presents the model not so much as a simple chatbot, but as a project assistant or an agent capable of planning and sequencing tasks on large data sets.
In Europe and Spain, this type of context is particularly interesting for sectors such as legal, financial, consulting, public administration or healthcarewhere it is common to handle very extensive documentation. The challenge will be to verify if the model maintains that capacity with reasonable latencies and a truly competitive cost per million tokens in production environments.
Reasoning abilities and agents: the bet to compete with closed systems
DeepSeek claims that V4-Pro It outperforms most current open-top models. In tests of global knowledge and reasoning, it only slightly lags behind some of the best closed-source models, such as the latest iterations of Gemini. In advanced reasoning, the company claims to be on par with premium solutions from OpenAI and Anthropic.
In the field of agents and linked tasksDeepSeek V4 has been designed to function as more than just a text generator. The company already uses V4-Pro internally for computer-assisted programming, workflow automation, and complex analysis, with the idea that the model can plan, execute and verify subtasks in a relatively autonomous way.
For Spanish companies that work with process automation, software development or data-intensive back officeThe combination of strong reasoning, long context, and low cost can be attractive. However, the actual maturity level of V4 agents and their robustness in Spanish These are still aspects to be evaluated in independent tests.
Open source versus closed models: impact for Europe
The V4 series is presented as open model with early accessDeepSeek has released V4-Pro and V4-Flash on platforms like Hugging Face in preview mode, allowing testing both in its own chat and via API. Some variants mention a MIT-type license which would open the door to local downloads and deployments without the usual restrictions of many business models.
This approach directly confronts the strategy of closed Silicon Valley modelsAnd with the ongoing debates about how to close open source, where access is often tied to API contracts, specific cloud providers, and higher prices, if DeepSeek V4 confirms that the performance gap between open and closed source has narrowed to almost nothing, many European companies might reconsider their reliance on US vendors.
In the EU regulatory context, marked by the AI Act and strict requirements regarding data sovereigntyHaving powerful models that can run on-premises or in European clouds without losing control over information becomes a key factor. However, Chinese open source also raises political and trust questions that governments and large corporations will need to carefully consider.
Alliance with Huawei and alternative hardware to NVIDIA
One of the most delicate elements of DeepSeek V4 is the change in the hardware baseFollowing controversies over the use of NVIDIA H800 GPUs and even accusations about the use of banned Blackwell chips, the company has strengthened ties with Huawei to reduce its exposure to US sanctions.
Shortly after announcing V4, Huawei officially announced that it will provide its Ascend chips and supernode systems to perform DeepSeek's inference tasks. Among them, the accelerators stand out. Atlas 350 powered by Ascend 950PR processorswhich have been specifically optimized for this type of MoE model and are already being sold as a local alternative to NVIDIA-based infrastructure.
The V4 technical documentation mentions the development of kernels adapted for both Huawei hardware and NVIDIA GPUsThis dual compatibility would allow the company to survive potential Western blockades while simultaneously leveraging the strength of the Chinese chip ecosystem. For Europe, this move opens the door to a powerful AI offering not entirely tied to the North American supply chain, although the use of hardware from China will also be scrutinized from the perspective of... cybersecurity and strategic dependence.
Pricing strategy: the hit to costs per million tokens
DeepSeek has long positioned itself as the "cheap reasoning" optionIt already did so with its R1 model and now reinforces the commitment with V4. In estimates advanced by the company itself and by API providers who are already testing the model, the cost of inference could be around $0,30 per million tokens, well below what many premium Western services charge.
Furthermore, DeepSeek maintains that, for the Flash variant, It will maintain similar rates to the V2 model. Launched in 2024, and has even hinted that prices could drop even further in the second half of 2026...as Huawei's Ascend 950PR supernodes are deployed on a large scale. This announcement has had immediate effects on the Chinese stock market, boosting local semiconductor companies.
In practical terms, for a Spanish company that processes lengthy legal documents, financial histories, or large volumes of customer interactions, this pricing structure allows it to operate with much higher margins to those obtained with more expensive APIs. However, it remains to be seen whether these rates will hold once V4 exits preview mode and whether some of the savings will be offset by greater implementation complexity compared to more mature services.
Availability, development status and access from Europe
DeepSeek V4 is, according to the company itself, in "preview" mode or early accessThe V4-Pro variant and its lighter version can already be tested through the official DeepSeek chat and via API, while some third-party providers, such as third-party platforms that offer test nodesThey have begun to give limited access to developers.
The company's roadmap has suffered certain delays compared to the initially leaked dateswhich pointed to full releases throughout the first half of 2026. Even so, in practice there are already V4 models posted in public repositories, ready to be tested and deployed in laboratory environments.
For European and Spanish organizations, access is currently focused on Public APIs, downloads from open source repositories, and testing on international cloudsDeployment on in-house infrastructure will depend on the ability to have compatible GPUs or, if necessary, approved Chinese hardware, something that may clash with internal security and regulatory compliance requirements.
Implications for startups and companies in Spain and Europe
For the entrepreneurial fabric of Spain and other EU countries, DeepSeek V4 represents a possible democratization of access to "enterprise" scale modelsWith much lower costs per million tokens and the option of local execution, projects that previously needed considerable funding rounds to cover expenses on APIs and GPUs can now consider high-level prototypes with more limited resources.
In sectors such as fintech, legaltech, digital health, or data analyticsThe ability to process enormous contexts and maintain data within infrastructures controlled by the company itself can make all the difference when it comes to meeting the European data protection regulationsHowever, the predominance of English and Chinese in official documentation, as well as the priority given to these languages in training, suggests that the Spanish requires additional fine-tuning work to achieve truly competitive performance.
Another issue is the maturity of the tools ecosystem around DeepSeek V4. Unlike models like Llama or GPT, which have open source frameworksVersion 4 starts with less out-of-the-box support. Companies with small technical teams will need to assess whether they can handle that integration curve or if they prefer to wait for the environment to become more established.
Geopolitical and regulatory dimension: AI as a new field of friction
The emergence of V4 occurs in a context in which AI is a vector of economic and strategic power.An open and competitive Chinese model puts pressure on the Western narrative of technological leadership and adds tension to the discussion about technological dependencies and global standards.
While China responds to the challenge with a industrial capacity offering, proprietary chips and rapid deploymentEurope tends to strengthen the regulatory component and the requirements for transparency, security, and ethics. The possibility that open models of Chinese origin could become global benchmarks may encourage European governments to raise regulatory firewallsFor example, in public procurement or critical sectors.
This could lead to a AI market fragmented by blocksIn this context, Spanish and European companies must choose between US models heavily constrained by their own rules of use, European solutions still under development, and powerful Chinese proposals with a complex geopolitical component. DeepSeek V4, with its combination of open source, low cost, and Huawei support, sits right at the heart of this debate.
Overall, DeepSeek V4 is shaping up to be one of the most serious contenders for bringing high-level language models into the realm of open source and aggressive pricingIts enormous context window, the alliance with alternative hardware to NVIDIA, and the focus on agents and reasoning make it an option to watch very closely from Spain and the rest of Europe; the real impact will depend on how it performs in production, the actual support in Spanish, and to what extent companies are willing to incorporate a key piece of their AI infrastructure from the Chinese technology ecosystem.
