We offer you a dive this week in the world of the six new Chinese Dragons of AI. And to start this series, it is difficult not to start with the actor who took everyone short a few months ago: Deepseek.
TL; Dr – Deepseek, the Chinese laboratory which forces Silicon Valley to revise its certainties
👥 For whom is it important?
- AI decision -makers and innovation managers in large groups
- Investors and analysts according to Sino-Western technological competition
- Product and R&D teams working on open source or low compute models
- Politicians and strategists in digital sovereignty
💡 Why is it strategic?
- Deepseek R1-0528 reaches GPT-4 performance with 5 % of estimated costs
- China shows that it can innovate outside the Californian frame: more sober, faster
- The use of “sparse” models opens the way to a frugal and distributed AI
- US export restrictions have accelerated Chinese technological autonomy
- Open Source becomes a lever for influence beyond speeches
🔧 What it changes concretely
- Competitive Open Source models, usable on GPU consumer public
- A new standard of cost/efficiency in AI for startups and laboratories
- Growing uses outside China (Perplexity AI, Hugging Face) despite tensions
- More lively debates on the sovereignty of the datasets and the regulatory biases
- Reinforced multipolar dynamic in the foundational AI
In the space of a few months, this young Chinese company, still largely unknown to the general public in early 2024, emerged in discussions around the most advanced artificial intelligence models. Deepseek, through the publication of its R1 model, has risen in technical comparisons with heavy goods vehicles like Openai or Anthropic. Without a crash, this breakthrough raises questions about the current dynamics of innovation, the side effects of technological regulations, and the role of China in the future balances of AI.
A remarkable performance, but still to be located
When Deepseek published R1 in early 2025, several observers noted its promising results on benchmarks of reasoning like MMLU or Arc. The model seemed to be achieved or even exceeding, certain GPT-4 performance, with a significantly lower calculation imprint. These results, disseminated in open source, were quickly taken up in third -party and community testing tools.
A few days ago, Deepseek published a major update of his model, called R1-0528. According to the first returns, it reduces the gap with commercial leaders such as Openai O3 or Gemini 2.5 Pro, in particular on the tasks of mathematical reasoning, generation of code and business logic. On the evaluation likes 2025, the success rate increased from 70 % to 87.5 %. Similar progress is observed on other technical benchmarks. The model remains accessible free of charge under MIT license, with downloadable weights from Hugging Face and a low -cost price API. These choices strengthen the dissemination of the model in the developer community, while consolidating the Deepseek position as a supplier of open source solutions at high level of performance.
A partial visibility strategy
Deepseek claims an “open” approach, at least in terms of publications and sharing certain weights of models. However, very little information is available on its governance, its sources of financing or the exact origin of its material infrastructure. Its founder, Liang Wenfeng, has not expressed itself publicly for over a year. No official spokesperson seems to be mandated to represent the company on the international scene.
Behind this discreet trajectory is an equally elusive character, Liang Wenfeng. Former quantitative trader, founder of High-Flyer Capital, Liang is described by his collaborators as meticulous, silent, but technically formidable. Little publicized, some in the sector nicknamed it “technical Madman”, he embodies another figure of leadership, more turned towards execution and architecture than to the story.
This discretion contrasts with the communication standards of the major American laboratories, which value relative transparency around their technical roadmaps or their organizational culture. The choice of Deepseek, whether tactical or cultural, gives to read another posture: that of an actor who favors execution to media coverage.
Constraints that have become lever
One of the specificities of Deepseek is due to its development context. Due to the American restrictions on the export of advanced GPU, the company would have designed its model around more sober architectures in calculation. Certain analyzes evoke an early recourse to “sparse” structures, optimized to take advantage of limited resources.
The release of a distilled model, Deepseek-R1-0528-QWEN3-8B, completes this logic. Lighter, this model makes it possible to perform complex tasks on standard GPUs type RTX 3060 or 4090. It aims for an audience of researchers, independent developers and limited medium companies, wishing to exploit recent advances without depending on a heavy infrastructure.
A collective dynamic to observe
Deepseek does not evolve alone. It is part of a new generation of Chinese actors, often locally nicknamed the “six small dragons”. In Hangzhou, where robotics startups, gaming champions doped at AI and companies from elite laboratories, is based on three pillars: local training in excellence, industrial density, and more explicit political support since 2022. Where Beijing firmly supervised its technological giants in 2021, the state seems now mobilized to strengthen industrial autonomy, in particular in artificial intelligence and semiconductors.
This dynamic is based on a return of profiles trained abroad, a demanding technical culture, and an ability to bring out very young teams on complex subjects. If this trajectory continues, it could strengthen a form of multipolarity in advances in artificial intelligence, beyond the current American-American duopoly.
Contrasted reactions and persistent controversy
In parallel with technical advances, Deepseek arouses strong geopolitical tensions. In the United States, parliamentary reports accuse him of having circumvented restrictions on the export of H100 chips, or even having accessed protected data. The company denies, without however providing a detailed response. This opacity strengthens questions, Deepseek wants to be open to its weights of models, but remains closed to its development conditions. This paradox fuels the distrust of part of the Western ecosystem, even in those who already integrate its models, often reconditioned, as with Perplexity AI, to respond to their own standards.
Reception in the technical community is generally positive. Several influential developers have praised the model’s ability to produce functional and coherent code from the first iterations, evoking performance comparable to those of paid models. This community recognition, if it continues, could weigh in the way the ecosystem now assesses Chinese contributions to the open source AI field.
A laboratory more than an upheaval?
What the Deepseek case reveals, beyond suspicion, is perhaps another path to IA performance, a localized ecosystem, sober in capital, rapid in its cycles, and based on a balance between engineering culture and values system partially offset from that of Silicon Valley.
📚 References and sources consulted
-
Deepseek ai. Presentation of the Deepseek-R1-0528 modelHugging Face.
https://huggingface.co/deepseek-ai/deepseek-r1-0528 -
Deepseek. API documentation – Pricing and integration.
https://api-docs.deepseek.com/quick_start/pricing -
Venturebeat (2025). Deepseek-R1-0528 Arrivals in Powerful Open-Source Challenge to Openai O3 and Gemini 2.5 Pro.
https://venturebeat.com/ai/deepseek-r1-0528-arrives-in-powerful-open-source-challenge-to-openai-o3-and-google-gemini-2-5-pro/ -
Reuters (2025). China’s Deepseek releases an update to its R1 Reasoning Model.
https://www.reuters.com/world/china/chinas-deepseek-releasses-an-update-its-r1-reasoning-model-2025-05-29/ -
The Guardian (2025). Ai Momer Likened to Sputnik as China’s Deepseek Shocks Western Labs.
https://www.theguardian.com/technology/2025/jan/28/deepseek-openai-elon-musk-don-lemon-crypto-kalshi -
Bloomberg businessweek (June 2025). The Architect of China’s Ai Revolution.
(Complete number, printed or digital version via Bloomberg.com) -
Reddit. Deepseek-R1-0528 Official Benchmarks released-discussion Thread on R/Localllama.
https://www.reddit.com/r/localllama/comments/1ky8vlm/deepseekr10528_official_benchmarks_released/ -
Elek Chen (2025). Deepseek-R1-0528 vs Claude 4-Full Comparison ReportMedium.
https://medium.com/@elekchen/deepseek-r1-0528-vs-claude-4-full-comparison-report-fefd9723D6FA
