A little-known Chinese AI research lab called DeepSeek has burst onto the global stage with the release of its open-source large language model, DeepSeek-R1, claiming it rivals—and in some cases surpasses—offerings from industry heavyweights like OpenAI. This move signals a potential shift in the artificial intelligence landscape, as DeepSeek proves that cutting-edge models can be developed with fewer computing resources than previously thought necessary.

DeepSeek’s origins trace back to Fire-Flyer, a deep-learning research branch of the Chinese quantitative hedge fund High-Flyer. Founded in 2015, High-Flyer quickly gained renown in China for its math- and AI-driven investment strategies, at one point managing more than 100 billion RMB (about USD 15 billion). However, in 2023, the firm’s co-founder, Liang Wenfeng, pivoted from purely financial pursuits to scientific exploration. He spun off Fire-Flyer into a standalone AI research company—DeepSeek—focused on creating next-generation AI models.

Unlike many Chinese AI ventures, DeepSeek has no financial backing from major tech companies such as Baidu or Alibaba. Liang has repeatedly stated his main motivation is scientific curiosity rather than immediate commercial gain. “Basic science research rarely offers high returns on investment,” he explained in an interview with Chinese publication 36Kr.

At the heart of the hype is DeepSeek-R1, an advanced reasoning model that claims to outperform leading models on a variety of math and logic benchmarks. Notably, DeepSeek has released not just its flagship model but also six smaller distilled variants under an MIT license—making them freely available for research, commercialisation, and further development.

Among these variations is DeepSeek-R1-Zero, which the company says was trained using only large-scale reinforcement learning (RL), emerging with robust reasoning abilities without the traditional supervised fine-tuning phase. DeepSeek-R1, on the other hand, refines these abilities and matches OpenAI’s o1 model in performance on reasoning tasks, according to internal benchmarks.

DeepSeek’s success takes on particular significance given the ongoing tech rivalry between the US and China, which has led to restrictions on exporting advanced chips—such as Nvidia’s H100—to Chinese AI firms. While High-Flyer’s early stockpile of around 10,000 H100s gave DeepSeek a head start, it soon became clear that scaling in the “Western way” (i.e., buying more and more top-tier hardware) would be difficult.

Forced by circumstance, DeepSeek turned to software-driven optimisations and innovative training strategies. These include:

• Multi-head Latent Attention (MLA) and Mixture-of-Experts: Techniques that split complex tasks among smaller, specialised sub-models, drastically reducing compute requirements.

• Custom Communication Schemes: More efficient data exchanges between chips to lower memory usage.

• Memory Optimisation: Reducing field sizes and fine-tuning architecture to maximise efficiency.

• Mix-of-Models Approach: Combining multiple smaller models to achieve performance levels comparable to, or even exceeding, those of much larger systems.

As a result, DeepSeek’s latest model reportedly required one-tenth the computing power of Meta’s comparable Llama 3.1 model to train, according to research from Epoch AI.

DeepSeek’s workforce consists largely of fresh graduates from China’s prestigious Peking University and Tsinghua University. Though they often lack industry experience, they bring academic rigor and a collaborative mindset that founder Liang says is perfect for tackling “high-investment, low-profit” research. Experts believe that many of these young researchers view advancing AI as a national mission—especially as the US-China tech standoff continues to limit their access to cutting-edge hardware.

Despite—or perhaps because of—ongoing US export controls, DeepSeek is forging a path that focuses on software-led innovation and efficiency gains. The company’s breakthrough underscores that while advanced chips are valuable, they are not the sole key to AI progress. It remains to be seen whether other Chinese AI firms will follow suit by doubling down on open-source research and computational optimisations.

For now, DeepSeek’s remarkable ascent stands as a testament to how necessity can drive invention—and how a small team of ambitious PhDs and a visionary founder can challenge the status quo in the race for AI supremacy.