DeepSeek and others

AI, China changes pace: no longer models but platforms for the real economy

From DeepSeek to Alibaba, Tencent and ByteDance, the challenge shifts from benchmarks to integration into business processes. Comprehensive architectures, competitive costs and increasingly agent-oriented models redefine the global race for artificial intelligence

by Alessandro Longo

10 May 2026

3' min read

Translated by AI

Versione italiana

3' min read

Translated by AI

Versione italiana

In recent months, if you said 'AI and China' only one name came to mind, at least to us Westerners: DeepSeek. The model that, for a moment, made OpenAI and Anthropic tremble. Wrong: the Chinese generative AI market is now a competitive ecosystem in which at least four families of models are going their separate ways: in addition to DeepSeek, which has just reached V4, there are Qwen3, Tencent's Hy3 preview (these two updated in April) and ByteDance's Seed2.0.

Another mistake: seeing this match only as a performance comparison between individual models. Like: is Gpt or Deepseek more capable, in (theoretical) benchmarks? In reality, China, just like the US, is already taking the next step. What we need if AI is to become the new operating system of our economy.

That is: to build a complete architecture, for an artificial intelligence that (increasingly, increasingly better) must be able to integrate itself into real work, at several levels, even autonomously. With different tools and cost control. Element, the latter, where China and its open models actually have a competitive advantage.

DeepSeek - this yes - remains an excellent case. It is now refining the formula that surprised the world in 2025. The one made of advanced reasoning capabilities, open licensing and very aggressive bee prices. With DeepSeek-V4, published on 24 April 2026, it wants to stop being just 'low-cost reasoning' and become a more comprehensive frontier model.

For an enterprise, DeepSeek is above all the way to flexibility. It can be attractive when the goal is to have more control over the model, to adapt it to specific contexts, to experiment with lighter versions or to contain the cost of using it on large volumes. The limitation is equally clear: an open and cost-effective model is not enough on its own. Turning it into value requires skills, control tools, integration with existing systems and solid governance.

Qwen from Alibaba, which reached 3.6 a few days ago, follows a different logic. It does not want to be just an efficient model: it wants to become a working base for those who develop applications, digital services and enterprise solutions. It is a strategy consistent with Alibaba's tech giant profile, which can connect AI, cloud, developer tools, business channels and enterprise customers.

Companies can choose lighter or more powerful models, balancing quality of response, speed and cost. Not all processes require the most sophisticated model. Some activities need speed and low cost; others require more reasoning power. Alibaba aims to cover this variety with a family of models, not with a single product.

Another historical Chinese tech giant is in the running: Tencent.

With Hy3 preview, presented at the end of April, it has an even different positioning. Its hallmark is integration in services and workflows. Tencent emphasises activities such as document analysis, information search, data processing, use of external tools and automation of complex steps.

It is an approach closer to the operational needs of companies. The model is presented not just as a machine capable of responding better, but as an engine to be inserted into products that are already in use: documents, browsers, assistants, productivity tools and internal applications. In this scheme, AI is not the final product, but a component working behind the scenes.

Finally, ByteDance. Known to us as the owner of TikTok. Its Seed2.0 aims to link AI research and consumer products with a very broad user base. ByteDance can test, distribute and improve its models in real, high-volume user environments.

This approach distinguishes Seed2.0 from other cases. The model is presented as a general purpose agent-oriented family, i.e. systems capable of dealing with complex, multi-step tasks. Not only chat, but AI capable of reading long documents, interpreting multimodal content, handling complex instructions and supporting high-value tasks.

China has many souls, we Westerners struggle to understand it. This also applies to AI. The competition between the US and China, with Europe far behind, is also played out on these complexities. Differences in approach, however, which must also be of interest to Italian companies. Understanding them allows us to understand how to derive more business value from the adoption of AI.