Latest News
Wednesday 29th January
Chinese artificial intelligence company DeepSeek has caused shockwaves across the tech industry and global stock markets this week after launching their latest AI models. The company reported that training their DeepSeek-V3 model, said to be on par with or superior to existing leading models, required a fraction of the computing power and costs standard in the sector.
The history of DeepSeek
The company was founded in May 2023 by Liang Wenfeng, operating independently with funding by High-Flyer, a hedge fund also founded by Wenfeng. Their first AI model, DeepSeek Coder, was released in November 2023, followed by DeepSeek LLM, DeepSeek-V2, and DeepSeek-Coder-V2. Their latest releases in January 2025, which have triggered the current market shock, are DeepSeek-V3 and DeepSeek-R1.
What makes DeepSeek’s AI models different from others?
A major differentiation between DeepSeek and competitors, including OpenAI and Meta, is their open-source approach. Anyone can download the DeepSeek-R1 model for free and run it locally, independently from DeepSeek. This open-source approach is expected to lead to further innovations across the tech sector, as any individual or business can adopt their model.
Additionally, DeepSeek utilises various innovative techniques, setting them apart from the rest of the market.
Traditional machine learning methods have relied on supervised fine-tuning, whereas DeepSeek uses a reinforcement learning model. This means that their models learn through trial and error and self-improve based on their results, leading to more sophisticated reasoning capabilities and adaptability.
DeepSeek’s models also use a different architecture to many other models on the market, known as a Mixture-of-experts (MoEs) architecture. A MoEs architecture only activates a small fraction of their parameters for a task, reducing computational costs and improving efficiency. While DeepSeek did not invent MoEs architecture (it was initially introduced in 1991), they have successfully used the approach innovatively to disrupt the AI landscape.
DeepSeek-V3 also utilises multi-head latent action, allowing the model to simultaneously handle multiple inputs, and distillation techniques allowing larger models to transfer knowledge and capabilities into smaller, more efficient models.
What has the effect on the market been?
Since DeepSeek released their new models and announced their significant computational and financial efficiency, the impact on the technology sector and stock market has been vast. Nvidia, who produce high-performance GPUs widely used for AI computing, were particularly affected, as DeepSeek’s model requires significantly less GPU resources to produce comparable outcomes to other existing models. On Monday 27th January, Nvidia experienced the largest market value drop in U.S. stock market history, losing USD 600 billion in value. The impact was felt widely, with the tech-weighted Nasdaq index dropping 3%. As of Wednesday 29th January, the market has partly recovered, but the impact is still being felt.
