Zixuan Li

At Z.ai, Zixuan Li’s work combined AI model capabilities, product design and distribution. In November 2025, ChinaTalk identified him as Z.ai’s Director of Product and genAI Strategy. His responsibilities included global partnerships, chat-model evaluation, API services and the GLM Coding Plan, the coding subscription he managed.[2]

Li described a shift from optimizing chat preferences toward useful work. In 2024, he said, Z.ai concentrated on Chatbot Arena and the answers users preferred. The arrival of Manus and Claude Code in 2025 changed the team’s priorities: he saw coding and agentic functions as more economically useful and capable of improving efficiency. “We are no longer putting simple chat at the top of our priorities,” he told ChinaTalk.[2]

The GLM Coding Plan addressed a product problem created by making flagship models open: customers could deploy the models themselves or buy access from another provider. Li argued that Z.ai needed to offer better engineering and faster decoding, alongside services such as search and MCP. He saw subscriptions as a way to retain users while relieving them of the need to track token consumption during coding interactions. This was his rationale for the product, rather than a measured retention result.[2]

Li framed openness as both a research contribution and a route to international adoption. Overseas customers, he explained, might use models through Fireworks or Groq, or deploy them on their own hardware, instead of using Z.ai’s API. He said DeepSeek R1 had demonstrated that releasing a model openly could bring recognition and commercial opportunities through APIs and collaborations. His formulation was: “You need to expand the cake first and then take a bite of it.”[2]

Li supported rapid release but wanted more time for partner preparation. In the November 2025 interview, he described the company’s practice as evaluating a finished model and open-sourcing it within hours. His own preference was a week to arrange trials with inference providers, benchmark companies and coding-agent developers before release. Instead, he sometimes had to negotiate integrations with only two or three hours’ notice.[2]

That coordination included discussions with Fireworks, LM Arena and Kilo Code. Li described meetings at 2 or 3 a.m. to accommodate US partners. He distinguished this schedule from research work: in his view, researchers needed rest to read papers, run experiments and write code effectively, rather than being expected to work around the clock.[2]

Li resisted assuming that more agents automatically meant a better product. In November 2025, he described Z.ai chat as a single actor that could search repeatedly and produce slides, presentations or posters. A single agent could see the conversation context directly; distributing work among agents required context compression that could lose information. He called for evaluating speed and performance across architectures before choosing a multi-agent approach.[2]

He also distinguished advertised context capacity from effective context. A nominal million-token window, he cautioned, might perform well only within 60,000 or 100,000 tokens; these were his illustrative figures, not independently verified benchmark results. He discussed compressing context to smaller windows as a practical engineering response, while arguing that some limits required architectural changes rather than more data alone. He described testing hypotheses on smaller models instead of the full 355-billion-parameter model.[2]

At the 2026 AI Engineer World’s Fair, Li presented virtually after travel issues prevented an in-person appearance. Richard MacManus’s July 1 account described his presentation of GLM-5.2 as Z.ai’s “flagship model for long-horizon tasks.” Li also introduced ZCode, a harness described as supporting all frontier models, and compared it with OpenAI’s Codex.[1]

Before these Z.ai responsibilities, Li had experience in both AI research and product work. He told ChinaTalk that he studied AI for science and AI safety at MIT, researching AI applications and alignment. His Conffab biography places that research before Z.ai and lists earlier roles as a data scientist at Apple and product director at a legal technology startup. It also lists master’s degrees in Information Technology from Carnegie Mellon University and Management Studies from MIT, without dates for those degrees or positions.[2][3]