The Chinese company DeepSeek has made another move in the global race for artificial intelligence when announcing DeepSeek-V3.2 and its variant V3.2-SpecialeThese two open-source models are aimed squarely at the high end of the market. The company claims its reasoning system is comparable to leading benchmarks like the GPT-5 and Gemini-3 Pro, putting pressure on the American giants in a time of intense technological competition.
In Europe, where debates on Responsible AI, regulation and technological sovereignty These trends are commonplace, and DeepSeek's move hasn't gone unnoticed. The fact that a Chinese lab has published weights, detailed technical documentation, and an advanced reasoning model in open source reinforces the feeling that the open-source ecosystem is regaining strength against completely proprietary solutions, something that could be particularly interesting for European universities, research centers, and tech SMEs.
DeepSeek-V3.2: reasoning at the level of leading models
The Hangzhou-based startup has presented DeepSeek-V3.2 as the final and stable version of its line of reasoning models, replacing the experimental edition released weeks earlier. According to the company itself, V3.2 achieves a performance similar to that of GPT-5 in various benchmarks audiences of multi-stage reasoning and thinking, and is positioned slightly below Gemini-3.0 Pro in some benchmark tests.
This model combines Human-type reasoning with the ability to use external toolssuch as web search engines, calculators, code execution environments, or third-party systems like Claude Code. The idea is that the system not only generates text, but can also plan, query resources, execute functions, and then integrate those results into a more complete response without requiring constant supervision.
DeepSeek has highlighted that the model offers two modes of interaction with toolsOne with visible reasoning, where the user can follow the intermediate steps, and another without showing the thought process. In both cases, the “Reasoning memory” persists amid calls to tools within the same conversation and restarts only when a new message arrives from the user, something especially useful for long tasks or agent-type flows.
The "thinking" mode integrated into the use of tools
One of the most striking new features of DeepSeek-V3.2 is the Direct integration of the thinking mode in the use of toolsWhile it reasons, the model can send queries to the search engine, invoke a calculator, execute code, or interact with other services, combining cycles of internal analysis and external calls to try to provide answers. more detailed and precise when the task requires it.
According to the company, this approach makes V3.2 its first model capable of reasoning and using tools nativelyboth in standard mode and in intensive thinking mode. It's a clear commitment to what are called agent-based workflowsIn these cases, AI does not simply answer a single question, but acts as an autonomous agent that breaks down the problem, searches for information, calculates, and then combines everything into a coherent solution.
DeepSeek also emphasizes that the model is widely available: DeepSeek-V3.2 can be used via web, app and APIThis facilitates its integration into products, virtual assistants, or business tools, including projects developed in Europe. For European developer communities and companies seeking open alternatives, the ability to explore and adapt the model without relying on a single major platform is a significant advantage.
DeepSeek Sparse Attention (DSA) architecture and computing efficiency
On a technical level, the core of DeepSeek-V3.2 is DeepSeek Sparse Attention (DSA), an attention mechanism designed to handle very long sequences while reducing computational cost. DeepSeek has unveiled a parallel file system optimized for AI which complements its efforts in efficiency and deployment. The model has around 671.000 billion total parametersbut at each inference step they are only activated around 37.000 billion parameters per tokenThis allows capacity to be maintained without increasing resource consumption.
This distributed architecture allows working with context windows of up to 128.000 tokens In production, this size is particularly useful for analyzing extensive documents, academic research, or reviewing large volumes of legal and technical information—areas of great interest to European institutions. According to data provided by the company, DSA reduces the cost of inference by approximately half in comparison to a previous dense architecture in long contexts.
For organizations in Spain and the rest of the EU facing computing budget constraints, this efficiency improvement It opens the door to experimenting with highly advanced models without needing the expensive infrastructure used by major US tech companies. Even so, DeepSeek acknowledges that it still has room for improvement compared to its competitors in token efficiency and breadth of world knowledge, two key areas for large-scale deployments.
DeepSeek-V3.2 with intensive reinforcement using RL and synthetic data for agents
Beyond the architecture, DeepSeek insists that much of the leap in reasoning comes from a massive post-training through reinforcement learning (RL)The company has allocated more than 10% of the total pre-workout calculation only at this phase, an unusual percentage in the sector, with the aim of strengthening the model's capacity to correct errors, reason in depth, use tools, and act in interactive environments.
The team has built a complex synthetic data ecosystem which includes more than 1.800 training environments and around 85.000 advanced instructions specific to agents. These tasks encompass real-world searches, dynamic simulations, code execution, chained problems, and automatically generated and verified scenarios to minimize errors in the dataset.
This approach is geared towards creating AI agents capable of operating with a degree of autonomyAnalyzing information, making decisions, and acting in multi-stage workflows. For European companies exploring the automation of complex processes—from financial analysis to advanced technical support—these advances may be particularly attractive, although it remains to be seen how the models will perform outside of controlled testing environments.
DeepSeek-V3.2-Speciale: mathematics, computer science and extended thinking
Alongside the generalist model, DeepSeek has launched DeepSeek-V3.2-Speciale, a variant geared towards advanced calculus, mathematical proofs, and extended thought processesThe company claims that this version is on par with Gemini-3 Pro Google's performance in complex reasoning tasks and that its performance approaches gold medal results in international competitions.
Specifically, Speciale would have reached levels comparable to gold medals in the International Mathematical Olympiad (IMO), International Olympiad in Informatics (IOI), ICPC World Finals and the Chinese Mathematical Olympiad. Furthermore, it integrates capabilities derived from the model DeepSeek-Math-V2, specializing in theorem proving and solving highly difficult problems, which strengthens its position in the segment of models for scientific and technical research.
Unlike the standard version, DeepSeek-V3.2-Speciale is not geared towards everyday tasks nor to generalist integrations with tools. The company emphasizes that this is a model designed primarily for research and academic work, with a consumption of Tokens superior, so for now It is only offered via API and not through general-purpose applications.
Availability of DeepSeek-V3.2, aperture and contrast with the American giants
DeepSeek has published the full DeepSeek-V3.2 weights and a detailed technical report regarding their training, something that contrasts with the increasingly restrictive policies of some large US tech companies, which often limit access to the code or the size of their most advanced models. Even in cases of open source Partial, like some versions of Llama, the opening comes with specific conditions and nuances.
In the European context, this degree of transparency and openness This can be key for projects that require auditability, regulatory compliance, or the ability to adapt models to regulatory frameworks such as the European Union AI ActUniversities, research centers, and public administrations can study the model in greater detail, replicate experiments, or even adjust some parts to their own needs without being completely dependent on a closed external API.
The company has put DeepSeek-V3.2 is available to the community on platforms such as Hugging Face and ModelScopeIn addition to offering access via API, the Speciale variant, on the other hand, is currently limited to consumption through a programmatic interface due to its higher computational demand and cost per tokenThis mixed distribution strategy fits with the interest of many European players in having robust models for research, although their commercial deployment may require more careful planning.
China's role in the global AI race
The release of DeepSeek-V3.2 comes at a time when China seeks to strengthen its leadership in AI Despite restrictions on access to advanced semiconductors and growing geopolitical tensions, DeepSeek has become one of the most talked-about names in the Chinese ecosystem after bursting onto the scene earlier this year with a model that surprised everyone with its power-to-cost ratio, and now it's doubling down with high-level agent and reasoning capabilities.
For Europe, where the discussion centers on how to balance innovation, data protection and securityThis type of development presents both opportunities and challenges. On the one hand, the existence of high-capacity open models from China expands the range of tools available to European laboratories and companies. On the other hand, questions arise regarding the compatibility with local regulations, cross-border data flows and the impact of content regulations in China, which some experts consider a possible obstacle to the full international expansion of these systems.
DeepSeek has also been gaining visibility outside its domestic market after its V3.1 model participated in automated investment experiments compared to systems like the GPT-5 and Gemini 2.5 Pro, where it showed competitive results. This strategy is complemented by the launch of other models such as DeepSeek-OCR, aimed at compressing text through visual perception and processing it with fewer resources, reinforcing the company's image as an actor focused on the efficiency and open source.
Expectations, limitations, and next steps
Despite the company's claims, DeepSeek acknowledges that V3.2 still lags behind some of its US counterparts in aspects such as general world knowledge, understanding of broad cultural contexts, or efficiency in token use. Furthermore, the project leaders themselves admit that the comparisons based on public benchmarks They do not always reflect real-world performance in production environments, especially in open tasks and with end users.
Another point to consider is that the integration of tools in reasoning mode It still needs to be thoroughly validated in complex, real-world use cases, from healthcare to financial or legal decision-making. The computational cost savings offered by DSA are significant, but they can be overshadowed if the quality of the responses is not consistently maintained when tasks become more ambiguous or require highly specific contexts.
With the arrival of DeepSeek-V3.2 and its Speciale variant, the landscape of advanced reasoning AI gains a new player with global aspirations, betting on open models, integrated tools, and contained costsThese developments broaden the range of options available in research, business, and public administrations, while forcing a rethink of how to fit the rapid evolution of AI into a demanding regulatory framework and increasingly marked competition between technological blocs.
