AI trends of 2025 – the innovation behind DeepSeek
In early 2025, DeepSeek suddenly emerged from deep Chinese waters with a revolutionary AI model: high-performance, open source, cheap and light. With that, DeepSeek entered the fray against US supremacy in AI, propelled by OpenAI, Microsoft, Google and Meta. Who will ultimately dominate the market remains uncertain, but no company – no matter how big – should rest on its laurels yet. What will drive AI innovation in 2025 and what brought us DeepSeek as a first surprise? We’ve identified three trends.
1. Open-source AI
Open-source software is software of which the source code is freely accessible. Developers can study, modify and improve it. Think of Linux or Android, software maintained and developed by thousands of people worldwide.
Open-source AI means that all essential components of the model are freely available: the training algorithms, the training data, the model parameters (weights), etc., so that others can reproduce, adapt and improve the model.
Yet the term ‘open source’ is often used more loosely. AI models, such as those from DeepSeek and from Meta, are more likely to fall under the term ‘open weights’: the model parameters are publicly available, while the training data and training algorithms remain secret. This makes these models more transparent, but they do not strictly meet the definition of ‘open source’. In practice, therefore, openness in AI (and software) is a spectrum rather than an absolute status.
“ Science and technology develop faster when a larger number of talented people are involved and share their innovations. ”
So, the line between open versus closed AI is open to debate, but proponents of open AI emphasise its benefits for innovation and progress. For example, Yann LeCun, Chief AI Scientist at Meta, states, ‘Science and technology develop faster when a larger number of talented people are involved and share their innovations’.
Opponents, however, warn of the risks: if advanced AI becomes freely accessible, it could also be used for harmful purposes, such as deepfakes, automated cyberattacks or misleading propaganda. At the same time, closed AI raises other problems, such as the danger of monopolisation, with only a few big tech companies controlling who has access to the AI system, who can develop it and what it can be used for.
DeepSeek’s models are a product of open-source innovation. The company builds on US technology, such as Meta’s LLaMA model and the PyTorch framework. But just as DeepSeek has benefited from open-source AI, it will in turn inspire others.
The real revolution is in the cost, which other companies are eager to emulate. DeepSeek trained its models with only a fraction of the resources that companies like OpenAI have required. This lowers the entry threshold for AI development and makes powerful AI more accessible to a much wider ecosystem.
2. Less computing power
DeepSeek performs at the same level as its US counterparts, but with much less computing power. According to estimates, a DeepSeek model consumes 20 to 40 times less energy than a comparable model by OpenAI – both in terms of training and usage.
Distillation
An initial explanation for DeepSeeks efficiency lies in a technique called distillation: a smaller model learns from a larger, more powerful model. OpenAI is currently investigating whether their models were used for this, which would be a possible violation of their terms of use. It would at least partly explain why DeepSeek could be built so cheaply: some of the development costs would have been borne indirectly by competitors.
Targeted calculations
DeepSeek's more efficient models (both V3 and R1) are not only the result of distillation, but equally the result of a series of other clever optimisations that drastically reduce the computational burden, as documented in the technical paper that appeared together with the model. (See box to the right)
The approaches discussed may seem obvious, but applying these techniques effectively requires strong engineering. And this is precisely where DeepSeek made a breakthrough. These modifications not only reduce the computing power needed, but also reduce the dependence on expensive hardware. For this reason, DeepSeek can also have an impact in the AI industry in the longer term.
Lower cost
If less computing power is needed to build advanced AI, development costs will also decrease. This could mean that AI development is no longer exclusive to billion-dollar companies such as OpenAI and Google, but that smaller and new players will get a chance to build competitive models and challenge the established ones.
Reduced ecological footprint
In addition, efficiency gains influence AI’s ecological footprint. Less computing power means less energy and water consumption.
Less Western dominance in AI
Lastly, these developments could cause geopolitical shifts. If China can build more efficient AI models without depending on Western dominance in hardware, it could fundamentally change the geopolitical balance of power in AI.
Targeted calculations
Instead of using brute computational power, targeted calculations are made to determine which element of the model ‘needs to work’ using a set of different techniques, which we compare below to the structure of a hospital.
Mixture of Experts (MoE)
First, it uses a Mixture of Experts (MoE) architecture, which means that only the most relevant parts of the model are activated when performing tasks. This saves computing power without sacrificing quality. It is similar to a hospital where a patient is sent directly to the right specialist instead of requiring a full team.
Multi-Head Latent Attention (MLA)
In addition, Multi-Head Latent Attention (MLA) further optimises the process by dynamically determining which parts of the input the model needs for a particular task. This avoids waste and makes calculations more targeted – such as a doctor immediately performing the most relevant tests instead of doing a general screening.
Whereas MoE distributes the workload among experts, MLA helps select the right information. Together, they make for more efficient AI that needs less computing power.
Quantisation
DeepSeek is also making strides when it comes to storage. Quantisation ensures that data is stored more compactly, requiring less hardware to provide the same storage capacity – comparable to smart packing so that everything fits into a smaller container.
Load balancing
Finally, load balancing takes care of balancing the distribution of work: it prevents some parts of the model from being overloaded while others remain unused. This acts like a hospital scheduler ensuring a fair distribution of patients among doctors.
Jevons paradox
More efficient AI may paradoxically stimulate more usage. When the threshold to develop and use AI lowers, the overall demand for computing power may increase. This phenomenon is known as Jevons paradox: the more efficient production is, the more consumption rises. The efficiency gains therefore do not lead to a decreasing need for resources, but rather to an increase.
3. New: large reasoning models (LRM)
In recent years, Large Language Models (LLMs) such as GPT-3 and GPT-4 have dominated the AI industry. But a new breed of models is due in 2025: Large Reasoning Models (LRMs). These models go beyond language understanding and generation and go through an explicit thought process before arriving at an answer.
The main difference between language and reasoning models is in the generation of the output and the transparency about it:
- Language models will mainly construct coherent answers based on text predictions without explaining exactly how the answers came about.
- Reasoning models take more ‘thinking time’ and consider multiple solution paths before formulating a conclusion. In addition, they apply ‘Multi-Step Reasoning’: complex questions are broken down into partial problems, which are analysed separately and then brought together. This ensures that, on average, they make more reliable and substantiated conclusions.
Deepseek vs OpenAI
Whereas language models often give a final answer without providing insight into their ‘thought process’, reasoning models make their reasoning visible. DeepSeek’s LRM (R1) incidentally goes even further than, for instance, OpenAI’s o1 model (also a reasoning model). The o1 model lists some intermediate steps but does not expose its entire ‘thought process’. This seems to be a deliberate choice by the developers, possibly for commercial or security reasons.
In contrast, R1 generates a fluid, natural line of reasoning, rather than a series of isolated reasoning steps. The model simulates a continuous line of reasoning, letting users literally follow along in the process.
A fluid line of reasoning has three major advantages:
- 1complex relationships and causal chains are worked out more explicitly,
- the model’s stream of ‘thoughts’ is easier for people to understand,
- users not only gain insight into the answer, but also the logic behind it.
Furthermore, Deepseek’s R1 model stands out in the way it presents its output. Whereas traditional AI models often generate long blocks of text, DeepSeek uses a schematic structure, which is better suited to systems thinking.
Transparency in AI
Many AI systems are like black boxes: they operate in a way that is opaque to both the providers of AI systems, the users who deploy AI systems, and the individuals affected by the use of AI systems.
Even with reasoning models, the mathematical calculations behind an answer remains elusive, but by making ‘reasoning steps’ explicit, reasoning models are more transparent than regular language models. Even though the underlying operation is still incomprehensible, the reasoning ensures that ordinary people have a better understanding of what is happening, allowing them to make targeted adjustments for better results.
Explainable AI
Reasoning AI is an important step towards so-called explainable AI. Together with extra ‘thinking time’ and Multi-Step Reasoning, reasoning models are particularly well suited for mathematical theorems, solving complex challenges with many dependencies, and for legal analysis, where context, transparency and accountability are crucial.
This is confirmed by large-scale benchmark research. When testing the capabilities of different models on specific problem types, LRMs consistently score higher than LLMs both for tasks requiring explicit multitask reasoning (MMLU) and for mathematical-based problem solving (MATH).
Reasoning models thus open the door to a new wave of AI applications. The shift from ‘text generation’ to ‘explicit reasoning’ could have a major impact in a variety of fields, from scientific research to strategic decision-making. Perhaps it will also make AI more human-centred, or contribute to AI transparency (a core element in the EU AI Act). Either way, DeepSeek has made an important contribution here in its own unique way.
Have you heard of these chatbots?
DeepSeek Chat
tool - online - DeepSeek
Le Chat
online chatbot - Mistral AI
Claude
chatbot - online - Anthropic
Gemini
tool - online - Google
Microsoft Copilot
chat tool - online - Microsoft
Jasper AI
copywriting tool - online - Jasper.AI
ChatGPT
Online chatbot - OpenAI
Johan Desseyn
Johan Desseyn is founder and researcher at Mpiris. He started Mpiris 2017 to help governments and organisations make good decisions. He does so by providing policy-oriented social science research. With Mpiris, Johan also responds fully to the digitalisation of our society. Where possible, he applies technology-based methods to answer his clients' questions.
For Johan and his team, the emergence of Generative AI meant not only looking for ways to use these tools internally in an effective and responsible way, but also committing to sharing the accumulated expertise with his clients in government and education.
Expertise: policy-oriented research, education and training, governments, labour market policy, social policy
Vinnie De Craim
I believe technology only truly matters when it creates societal impact. With a background in sociology and technology, I work on digital innovation in the public sector—where thinking and doing go hand in hand.
My approach is hands-on and impact-oriented. I bring critical analysis, but I’m equally driven by building concrete, meaningful results. Whether it's generative AI, policy development, or science communication, I’m always looking to bridge knowledge and real-world application—with pragmatism and a keen eye for opportunity.
I also believe in the power of language and storytelling to make complex ideas accessible. My background in the arts helps me not just understand technology and policy, but translate them in ways that resonate with different audiences.
Expertise: statistics, generative AI/AI, data science
Related courses
Machine learning and deep learning fundamentals
Workshop - Leuven - VIB, Training geselecteerd in de Belgische AI Factory Antenna (AIFA)
