Tsinghua's top graduate joins DeepSeek to tackle AI model efficiency

Gu Yuxian, a PhD graduate from Tsinghua University's computer science department and recipient of the prestigious 2025 Graduate Special Scholarship, has officially joined DeepSeek. His name already appears on the author list of the DeepSeek V4 paper.

Gu's research focuses on efficiency optimization for large language models during pretraining, compression, and inference. He has nearly 5,000 Google Scholar citations. His notable works include MiniLLM, a knowledge distillation method for large models that's been adopted by Google, Alibaba, and Nvidia, and the hybrid architecture model Jet-Nemotron. On an H100 GPU processing a 256K ultra-long context, Jet-Nemotron achieves up to 53.6 times the generation throughput of traditional full-attention models, and it outperforms larger mixture-of-experts models on multiple benchmarks.