Zeng Zhiyuan / 曾志远

Zeng Zhiyuan / 曾志远 appears in 177: 详解Kimi K3：强到冲击Anthropic估值的模型什么样？ as the algorithm-side guest, identified by the source as a University of Washington PhD student. He evaluates Kimi K3 through long-horizon agent behavior, frontend generation, architecture choices, expert routing, optimizer design, post-training, and distillation.

His source-specific view is that K3’s strengths are not a single breakthrough. They come from targeted evaluation/data loops, hybrid attention, NoPE, Attention Residues, Quantile Balancing, Per-Head Muon, and post-training workflows such as MOPD and On-Policy Distillation.

Connections

• Kimi K3, Kimi Linear, Kimi Delta Attention / KDA, and NoPE / No Position Encoding — model family and architecture context.
• Attention Residues, Quantile Balancing, Per-Head Muon, and Mixture of Experts — algorithm and training-stability branch.
• MOPD Post-Training, On-Policy Distillation, Model Distillation / 模型蒸馏, and AI Verification — post-training and reward branch.
• Recursive Self-Improvement, ML Coding, and Model Harness Co-Evolution — model-improvement loop context.