DeepSeek is collaborating with Tsinghua University to streamline the training process of its AI models, aiming to reduce operational expenses.
The Chinese startup, which made waves in the market with its low-cost reasoning model launched in January, partnered with researchers from the prestigious Beijing institution on a research paper presenting an innovative approach to reinforcement learning, enhancing model efficiency.
This novel method is designed to align artificial intelligence models more closely with human preferences by providing rewards for delivering accurate and easily understandable responses, according to the researchers. While reinforcement learning has shown success in accelerating AI tasks in specific contexts, its application to broader areas has faced obstacles. DeepSeek’s team aims to address this challenge through a technique it refers to as self-principled critique tuning. The strategy has demonstrated superior performance compared to existing methods and models across various benchmarks, achieving better results with reduced computational resources, as highlighted in the paper.
The new models developed by DeepSeek are termed DeepSeek-GRM, which stands for “generalist reward modeling,” and will be made available on an open-source basis, the company announced. Other AI developers, including the Chinese tech giant Alibaba Group Holding and San Francisco-based OpenAI, are also venturing into novel territories focused on enhancing reasoning and self-improving abilities of models during real-time task execution.
Over the weekend, Meta Platforms Inc., based in Menlo Park, California, unveiled its latest series of AI models known as Llama 4. This release marks the company’s first to implement the Mixture of Experts (MoE) architecture. DeepSeek’s models leverage MoE heavily to optimize resource utilization, and Meta has benchmarked its latest release against the startup from Hangzhou. However, DeepSeek has yet to announce a timeline for the launch of its next flagship model.
© 2025 Bloomberg LP
(This story has not been edited by NDTV staff and is auto-generated from a syndicated feed.)
