Zixuan Li leads Z.ai’s global ecosystem, with responsibility for its chat product, API services, international partnerships, and branding. His work connects the GLM family of foundation models to developers and businesses adapting them. His product strategy emphasizes sustained agent performance, specialized applications, and users’ control over how models are deployed.
From legal technology to AI products
Li’s earlier career crossed legal technology and data science. He worked in product leadership at iCourt and later in data science at Apple. His professional biography also describes research on AI alignment and generative AI applications at MIT before joining Z.ai.
His education spans Renmin University of China, Carnegie Mellon University, and Tsinghua University. He completed MIT Sloan’s Master of Science in Management Studies in 2024. At Z.ai, his responsibilities bring product development and international commercialization together: chat and API services give users access to the models, while partnerships support their use in other products and specialized workflows.
Li is among the co-authors of the GLM-5 technical report, a large collaborative effort focused on agentic engineering. The team’s work combines sparse attention with asynchronous reinforcement learning, separating the generation of training experiences from model updates to improve post-training efficiency. Li’s public arguments focus on what model advances should enable: agents that complete difficult work, improve an initial solution, and remain useful through repeated interactions.
Agents that improve and recover
Li distinguishes long-running execution from sustained improvement. Counting to a million may take hours, but it does not demonstrate that an agent can develop a better solution. His account of long-horizon work emphasizes revisiting an approach, testing alternatives, and turning additional effort into better results. Optimizing a CUDA kernel provides a concrete example: the useful outcome is improved performance, rather than the time spent working.
That distinction makes reliability a product concern as well as a model capability. Li argues that device agents must recover from mistakes and remember unfinished work across sessions. Small delays accumulate across interface actions, while a misclick can derail an otherwise successful workflow. In his discussion of device-agent reliability, he places particular weight on recovery: users may accept slower execution when an agent finishes dependably. Lost instructions after context compression and errors carried into later steps are obstacles to that dependability.
Li also treats user feedback and realistic evaluations as inputs to model development. Direct conversations help identify which capabilities need attention, while evaluations must reflect actual coding and agent workflows, including imperfect operating conditions. Requests for vision, speed, and longer context sit alongside the basic requirement that an agent solve the task well.
Open weights as a product strategy
Li advocates open weights because they let users adapt both the model and its deployment to their needs. In his GLM-5.2 keynote, he explains three complementary reasons for that approach:
Deployment control: Enterprises and governments can run models on their own infrastructure. Li connects that control to security needs and trust in the provider.
Specialized capabilities: Companies can fine-tune models for legal, financial, or security work. Li cites Harvey’s fine-tuning of GLM-4.1 as an example, describing adaptation as a way for application builders to differentiate their products.
Shared development: Access to model architecture and published training methods helps customers and developers investigate the technology and influence its direction. Li credits inference, fine-tuning, and application builders with contributing to GLM’s success.
The GLM-5.2 release makes the weights available under an MIT license and documents local serving options. Li’s position connects that openness to Z.ai’s business: downloadable models support independent deployment and adaptation, while managed services serve customers who want help using the technology.
His advocacy for ZCode extends the same practical focus to the software around the model. He introduced it as a coding harness built for GLM-5.2 that also supported other frontier models through users’ own API keys. ZCode’s Goals feature organizes continued planning, execution, and verification for complex development tasks. Li’s emphasis on iterative improvement therefore reaches beyond model training to the environment in which developers direct ongoing work.
Zixuan Li introduces Z.ai’s latest model through long-horizon coding, thinking budgets and general-purpose capabilities, then explains how open weights support local deployment, specialization and collaboration. The closing announcement, Z Code, adds a coding harness around the model.
GLM-5.2’s reported non-thinking improvement over GLM-5.1 with thinking separates gains in model capability from gains obtained by spending more thinking tokens.
Open weights support both local deployment and domain fine-tuning; architecture and training-recipe visibility support a further goal of collaborative model development.
▶ Watch ↗