Two Chinese Breakthroughs Revolutionize AI Model Manufacturing
DeepSeek and ModelBest are revolutionizing the AI industry by implementing innovative strategies to enhance model efficiency, focusing on advanced training techniques and densification.
Two Chinese Companies Revolutionize Model Manufacturing Approaches
In the expansive realm of artificial intelligence (AI), three groups of competitors stand out based on their approaches to large-scale model manufacturing. At the forefront of this race are two innovative Chinese startups, DeepSeek and ModelBest, whose methodologies starkly contrast with that of OpenAI.
Two Distinct Paths to Efficiency
While both companies aim to enhance the efficiency of large-scale models, they pursue markedly different strategies. DeepSeek focuses its innovations on the training phase, having developed a mixed experts model (MoE) from scratch and employed multi-head latent attention (MLA) technology.
“This solution optimizes both large cluster training and deployment services, leveraging cloud computing to provide services to users,” industry professionals at AItechtalk explain.
Conversely, ModelBest boosts efficiency during the inference (deployment) phase. Its MiniCPM model “enhances model training technology to increase the knowledge density of models, then challenges large-parameter models with fewer parameters.”
A New Law for Large Models
ModelBest’s founder, Zhiyuan Liu, introduces a Densification Law for large models. This law tracks the continual increase in the ‘density’ of large models’ capabilities, meaning achieving the same performance level with fewer parameters.
A Promising Future
Client-side computing power still poses a potential bottleneck. However, Han Xu, chief researcher at ModelBest, remains optimistic about future developments, noting that “the technologies are no longer problematic, it’s just a matter of waiting for the right opportunity.” The challenge ahead lies in optimizing the relationship between model training and chip development to capitalize on recent technological advances. For Han Xu, the future belongs to agile and optimized models.