DeepSeek V3: Open-Source Powerhouse

In this episode we take a close look at the DeepSeek V3 paper: the open-source powerhouse that is currently causing a stir in the AI scene. We explain why this model, with its 671 billion parameters (37 billion active per token) and innovative architectures such as Multi-Head Latent Attention and Mixture-of-Experts, is revolutionising the market. We examine how DeepSeek V3 achieves impressive results despite low training costs (only around 5.58 million USD) through efficient use of Nvidia H800 chips and lean data usage. Find out what lies behind the hype, what technical innovations distinguish the model, and why it is considered a game-changer in open-source AI.