陈广宇 Guangyu Chen

A machine learning researcher focused on LLM architecture and efficient attention. He joined Moonshot AI's Kimi team in November 2025 as an ML researcher, won the team's internal 48-hour hackathon shortly after, and entered the core research group. In March 2026 he published Attention Residuals as one of three equal-contribution first authors — rethinking the residual connection that Transformers had used for nearly a decade, cutting training compute to ~80% at equal quality. The technique later became a foundation of Kimi K3 (2.8T parameters). A USACO Platinum finalist and core contributor to the FLA open-source project (4.7k+ GitHub stars), his story has been covered by Xinhua-affiliated and major tech media.

• Proposed replacing fixed-weight residual accumulation in Transformers with learned, input-dependent attention over depth
• Co-designed Block AttnRes with Yu Zhang: partitioning ~100 layers into ~8 blocks, reducing memory from O(Ld) to O(Nd) and making the method practical at frontier scale
• Validated on Kimi Linear 48B (1.4T tokens pre-training): matching baseline quality with ~80% of the compute; GPQA-Diamond +7.5
• Shipped into production as an architectural foundation of Kimi K3 (July 2026, the first open-weight 3T-scale model)
• Open source: github.com/MoonshotAI/Attention-Residuals

⚠️ Note: behavioral inferences from public reporting and open-source records — not an official assessment by the individual or any organization.