Yuxuan Zhang is an algorithm engineer at Z.ai building the open-source GLM ecosystem, including language models, multimodal systems, and software agents. His work addresses the practical requirements behind capable AI: training models on real codebases, evaluating complete agent workflows, scaling reinforcement learning, and deploying large open models.
Earlier, Zhang earned national-level awards in RoboMaster and China’s National Students’ SmartCar Competition and contributed to Langchain-Chatchat and chatpdf. He is pursuing a PhD at the University of Liverpool and is a visiting PhD researcher at Mohamed bin Zayed University of Artificial Intelligence.
At Z.ai, his contributions span GLM-5, GLM-V, CogVideoX, CogAgent, and Open-AutoGLM. He is a named contributor to the GLM-5 technical report. At AI Engineer Code 2025, he described a GLM ecosystem that had surpassed 100 million downloads and detailed the engineering behind GLM-4.6.
- Evaluate coding agents as complete systems. Zhang highlighted CC-Bench, which assesses frontend development, internal tools, data analysis, and algorithms while recording planning, tool calls, code edits, and execution.
- Teach models to navigate actual repositories. GLM-4.6’s training builds from approximately 15 trillion general-pretraining tokens toward code and reasoning data, interconnected files, issues, pull requests, long-context examples, and complete agent trajectories.
- Match reinforcement-learning architecture to the workload. The open slime framework combines SGLang inference with Megatron-based training. Short reasoning tasks benefit from synchronized updates; asynchronous agent reinforcement learning lets slower environments feed a shared buffer without blocking training. Zhang also emphasizes curriculum difficulty, token-level optimization, and expert-verified data.
- Give multimodal agents usable real-world context. GLM-V preserves native image dimensions for screenshots and slides, while temporal index tokens help video models track event order. Open-AutoGLM extends these capabilities into phone-based interaction. In a Z.ai community discussion, Zhang also clarified that deploying a 355-billion-parameter model at FP8 requires substantial memory beyond model weights, including capacity for the key-value cache.
