According to Reuters reports, DeepSeek unveiled an update to its low-cost artificial intelligence model on March 25, 2025, increasing its competitive level against AI industry leaders such as ChatGPT producer OpenAI and Anthropic. The latest version of the offering, DeepSeek-V3-0324, has been made available open source through the AI development platform Hugging Face.
How do DeepSeek and Openai specifically differ in their approach to genarative A? How can you also benefit from it in your everyday work?
To find out, keep reading this article!
DeepSeek VS OpenAI
The tension between DeepSeek and OpenAI began in 2023, when the fledgling Chinese startup entered the scene with a clear message: generative artificial intelligence should not be the exclusive property of big tech. Within months, DeepSeek released DeepSeek-VL, an open source, free multimodal (text + images) model. As a result, while OpenAI continues to make its models accessible only through paid APIs and commercial licenses, DeepSeek takes the opposite route: open code, permissive licenses, full sharing.
In response to the Chinese company, OpenAI strengthens its market position through partnerships with giants such as Microsoft, maintaining a technical edge but moving further and further away from a collaborative vision. DeepSeek, however, gains attention and funding by proposing a participatory, distributed AI model that quickly becomes a viable alternative for developers and companies.
In recent days, the challenge has intensified: DeepSeek has released a substantial update, the DeepSeek-V3-0324 model, posted on Hugging Face with all the weights and code making it immediately accessible to developers, companies, and researchers worldwide. Performance is moving closer and closer to GPT-4, with major improvements in reasoning, image and text handling, and the ability to respond in English.
In an environment dominated by paid APIs and closed models such as those of OpenAI, this is a clean break. DeepSeek presents itself as a free alternative to ChatGPT, challenging not only on a technical level, but also on an ideological one: AI as an open infrastructure on which to build, collaborate, and innovate. The release has already had tangible effects on the market with declines in Nvidia shares and open discussions among tech CEOs such as Satya Nadella. This affair signals a structural change: China is not only participating in the global AI race, it is beginning to rewrite the rules.
Architectures compared: why DeepSeek and ChatGPT work differently
A crucial technical element that distinguishes DeepSeek from ChatGPT concerns the architecture of the underlying models.
ChatGPT, as Sam Altman explained to us at Italian Week Tech 2024, is based on the Transformer architecture, which has become the standard for natural language processing thanks to the self-attention mechanism, which allows the model to analyze all the words in a text simultaneously, attributing contextual weight to each. This structure enables deep language comprehension, which is ideal for producing coherent, fluid and natural texts. ChatGPT thus excels in generating conversational content, educational explanations, creative writing and generalist assistance. However, the entire model is triggered with each request, which incurs a high computational cost and can slow down responses in more complex tasks.
DeepSeek instead adopts a strategy based on Mixture-of-Experts (MoE): an architecture composed of specialized modules, each trained on different tasks. When the user sends input, only a targeted selection of these “experts” is triggered through an intelligent routing system. This approach allows the model to be more efficient, with lower resource consumption, and to provide more timely responses in highly specialized areas such as programming, mathematical analysis or financial modeling. The selectivity of the computation is one of the keys to optimization that makes DeepSeek particularly suitable for technical and data-intensive scenarios.
These architectural differences also reflect two visions of the technology: ChatGPT as a AI generalist oriented interaction, DeepSeek as a technical tool for precision and scalability. But these two paths, separate today, are already beginning to converge. Indeed, some recent developments point to hybrid models that combine the conversational fluidity of Transformers with the modular efficiency of MoE. It is in this direction that the future of generative AI is moving: systems capable of adapting to the context of use, efficient and accessible, but also flexible, dialogic, and transparent.
All this is also about Social Thingum
For those who, like us at Social Thingum, work with digital technologies for social impact, this is not just industry news. It is a matter of method, of accessibility, of vision. DeepSeek is proving that it is possible to create advanced and powerful models without hiding behind commercial licenses and walls of APIs. The Chinese company is hinting at something we have long repeated in our work: innovation can (and should) be shared. AI does not have to belong to a few, but can become a distributed tool, capable of generating value where it is most needed – even in fragile, local, community contexts.
In Social Thingum we develop digital technologies to make public and social services more understandable, accessible and participatory. We are a company that works with public agencies, foundations and third sector organizations. For years we have been experimenting with tools based on generative artificial intelligence not to replace people, but to help them:
- We have built interfaces and digital assistants designed for contexts where technological mediation is needed.
- We have built platforms for cultural inclusion where AI helps tell the story of local heritage in narrative form.
- We have designed automated synthesis systems to make regulations, notices, and complex information understandable.
- In every instance, we choose open source, transparent, interoperable tools. Because we believe in an AI that communities can understand, modify, adapt.
Conclusions
The game DeepSeek is playing against OpenAI is not just a competition between models. It is a clash of worldviews. Centralization versus distribution. Limited access versus openness. APIs versus sharing. And for those working in the field, in schools, in social services, this difference really matters because it makes us realize that another AI is possible. And that if we want to build it, we have to do it together-with open tools, adaptable models, and one question always on our minds: for whom are we designing these technologies?
We at Social Thingum will continue to work in this direction. Because AI cannot be a black box. It has to be a readable, controllable, human tool. And whenever we can, we choose openness as our starting point. This is also why we look to DeepSeek not only with technical curiosity, but with a certain attunement.
Want to know how we can help your company integrate these systems? Contact us!