The Z.ai Playbook

"The Z.ai Playbook" is a ChinaTalk podcast episode published on 21 November 2025, hosted by Jordan Schneider with Irene Zhang and Nathan Lambert as co-hosts (Lily Ottinger is credited on the post but does not speak). The guest is Zixuan Li, Director of Product and genAI Strategy at z-ai, also known as Zhipu AI. Li runs global partnerships, chat model evaluation, the API business and the GLM Coding Plan, and studied AI for science and AI safety at MIT. The conversation took place around the GLM-4.6 release, after GLM-4.5 had been called "another DeepSeek moment" by Lambert. In his open-source-ai-reading-list, Lambert quotes one line from it as evidence of how fast Chinese labs ship: "Get it out fast. We open source it within a few hours."

From AMiner to coding agents

Li says Zhipu was founded in 2019 to pursue AGI through graph computing rather than language models, and its early product was AMiner, a Google Scholar-like map of researchers and papers. It moved to LLMs in 2020 and published the GLM paper in 2021, which he places a year ahead of GPT-3.5. In 2024 the goal was chat quality, and GLM ranked roughly sixth to ninth on Chatbot Arena. In 2025 Manus and Claude Code convinced the company that coding and agentic work were worth more economically, and chat stopped being the top priority. Li says Z.ai became known through GLM-4.5 and 4.6 working inside tools like claude-code and Kilo Code: "We need to combine with these top products to gain fame."

How the lab works

Li describes the pre-training and fine-tuning teams sitting next to each other on one goal, a single model that does reasoning, agentic tool use and coding. For GLM-4.5 they trained three separate teacher models and distilled them into one, which he credits for building it efficiently (see llm-distillation). He insists team heads have to run experiments themselves, because a new Claude or GPT can land in the middle of a training run and you have to feel the shift directly; even the founder reads papers and runs experiments.

Many active PhD students work on GLM and treat it as their main achievement. Hiring follows a hierarchy. ByteDance and Alibaba pay more and get first pick, so startups like Z.ai recruit people who want to fight. The core research team is 100 to 200 people, similar to OpenAI by Lambert's estimate. At large companies Li thinks each team has 10 to 20 core members with 80 to 100 more on training and data preparation. On hours, Li works 18 a day because he negotiates with American partners on their time zone, sometimes at 2 or 3am, but says researchers get eight useful hours and should rest after that. Lambert jokes about Silicon Valley's "002".

Why a Chinese lab open-sources

Li gives two reasons in order. The first is contributing to research, as Llama, Qwen and Kimi do. The second is commercial. American companies will not call a Chinese API. They will run the model on Fireworks or Groq, or on their own hardware, and need the data stored in the US, so being open is the only way to be used there at all. Lambert confirms he could not sign up for a Chinese enterprise API at Ai2 but distills from several Chinese models. Z.ai's flagship was closed in 2024, and DeepSeek R1 changed that by showing a lab could become famous from open weights and still earn from its API. Li's summary: "You need to expand the cake first and then take a bite of it."

Fame in the US matters because the developer and research community there sets the agenda, and Chinese media follows American influencers, so Chinese enterprise buyers watch global reputation too. When Karpathy, Altman or Musk tweet about a model, both markets hear. Schneider calls the loop "recursive": Chinese media covering the Chinese models Americans talk about. Li thinks Z.ai still loses at branding. A German professor he met knew Llama, Qwen and Mistral but not GLM, Reddit threads still ask what GLM is, and Z.ai had 20,000 followers on X against roughly a million for DeepSeek.

Earning money from open weights

Li splits Chinese enterprise buyers in two. Some cannot send data to anyone, even Z.ai or Alibaba, and have integrators deploy DeepSeek on their own chips with RAG and workflows on top. Others, mostly tech and media companies, buy APIs on price and performance, a market he thinks ByteDance dominates. Qwen keeps Qwen 3 Max closed to have something to sell. Because Z.ai's flagship is open, customers ask why they should pay, so Z.ai competes with other GLM hosts on engineering: faster decoding, search, MCP.

Subscriptions turned out to work. The GLM Coding Plan spares users from counting tokens, and Li does not need a majority: if 5% of Claude users switch, "it's a huge market". Z.ai also sells into role-play, where GLM is popular on SillyTavern and Janitor AI. GLM-4.5 was weak at it until post-training data taught GLM-4.6 to hold a character through a long system prompt instead of forgetting who it is. His example is a custom Stewie from Family Guy.

Z.ai runs two platforms, BigModel.cn at home and Z.ai abroad, hosted in Singapore; Li is an employee of the Singapore company. Overseas inference runs outside China for data-residency reasons, and training stays in China because that is where the compute is. India has the most overseas users, but the US produces 50% of overseas revenue because Americans buy the Pro and Max plans. Growth came through X, Reddit and YouTube.

Translation, walled gardens and search

Li puts GLM's Chinese-English translation on par with Gemini 2.5 Pro and calls memes "one of our weapons". A whale emoji in a sentence about AI gets translated as DeepSeek, and in a sentence about animals as "whale". Walled gardens were not the obstacle Schneider had expected. Z.ai cannot scrape WeChat, so it generates synthetic data and studies public comment sections on Xiaohongshu and TikTok, collects danmu, and reads image memes with its vision model. The "TikTok refugees" episode helped by creating demand for auto-translation. GLM handles about 20 languages well. For search, Li says Google has no search API, Bing is withdrawing its own, and scraping resellers are fragile, so agents increasingly browse pages directly.

Chips, fear and the wall

Z.ai uses both Huawei and Nvidia chips and matches them to use cases by required speed, 30 tokens per second for one customer and 80 for another. On fear, Li says in China it belongs mainly to developers and data analysts, who use coding agents and see what they can do, and it is about jobs, not AI taking over. Nobody at Z.ai worries about jobs because the lab is short of people. Emotional attachment and "AI psychosis" are not a priority yet; Li argues the model has to reach GPT-5 level before de-addiction work is worth doing, since behavior changes with every version. Lambert contrasts this with the US, where a vocal minority fears much more than job loss, and Li remembers that everyone at MIT was pessimistic.

On research, Li says Z.ai is mature at off-policy RL and still exploring on-policy RL and multi-agent systems, which trade the single agent's full context for compressed per-agent context and can be ruined by one hallucinating agent. Models claiming a million-token window work well only to about 60k or 100k. He believes the current architecture has a wall that data alone cannot cross. GLM is 355 billion parameters, experiments run at 9B or 30B, and 90% of them fail. The next release was GLM-4.6 Air, about 30B (he was unsure whether it would be called Mini), and ideas tested in 2026 would reach production later. Later GLM generations are covered in glm-5-2-step-change-for-open-agents and glm-5-3-how-chinese-labs-keep-stride.

Z.ai releases within hours of finishing evaluation, with no pre-release LM Arena endpoint and no anonymous teaser campaign, because "the open source itself is the biggest event". Li would like a week to line up inference providers, benchmarkers and coding agents, and instead calls partners at night with two hours' warning. At home, even Tsinghua students have not heard of Z.ai, and high-schoolers ask why they need it when they have DeepSeek.

Against the other pages

Li's account of why Z.ai is open is the most commercial one in this group: a closed flagship in 2024, and open weights as the only way into the American market. That is narrower than the academic default in Kevin Xu's chinas-structural-advantage-in-open-source-ai and the two-decade open source culture in his chinese-open-source-a-definitive-history, though it matches Xu's point that open source is a precondition for growth abroad. Lambert's notes-from-inside-chinas-ai-labs, written after visiting Z.ai six months later, agrees on the student-heavy teams and the practical attitude to openness. Li's claim that Z.ai trains on synthetic data because it cannot scrape the walled gardens is the practical side of the data question Xu raises.

The transcript has machine-transcription errors. "Cloud Code" means Claude Code, "graphic network or graphic compute" most likely means graph networks, and "manuscript" in the search answer is probably a browsing agent's name misheard. A published correction fixed one misattribution (Kimi K2 for Manus and Claude Code). The model-naming answer about 4.6 Air and Mini is muddled in the source itself.