Hours after training ends, Z.ai puts the weights on the internet

Nathan Labenz opens this Cognitive Revolution cross-post by placing a lab most Western listeners still skip. As of the recording, Z.ai's GLM 4.6 sits 19th on the LMArena text leaderboard. Its Elo is roughly 65 points below the current leaders, which still means it wins two of five head-to-head comparisons. It happens to sit next to Qwen 3 Max, Kimi K2 Thinking, and DeepSeek v3.2. Those four, all from China, are the top open-weight models available, with Mistral not far behind.

On the Web Development board, GLM 4.6 does better: ninth, competitive with GPT-5.1 and meaningfully behind only Gemini 3 and Claude 4.5 Opus. The conversation itself is a ChinaTalk episode hosted by Jordan Schneider, with analyst Irene Zhang and Nathan Lambert of Ai2 and Interconnects. Their guest is Zixuan Li, who runs global partnerships, Z.ai chat, model evaluation, the API, and the GLM Coding Plan.

Labenz flags the themes that follow: open weights as a marketing tactic rather than an ideology, role-play as a Chinese use case that pulls fine-tuning in a different direction, Silicon Valley voices that confer credibility even inside China, a talent market where almost nobody at Z.ai knows Li studied at MIT, her view that "there is a wall," and a release cadence that can put weights on the internet hours after training ends. Chinese labs, he says, still talk about themselves as upstarts trying to keep pace, not as peers of the American frontier.

Li studied AI for science and AI safety at MIT, then came home because her family is in China. She applied to Moonshot, MiniMax, and others, and was rejected or ignored in a flood of résumés. A month after graduation she interviewed at Z.ai. Her first job was not overseas growth. The company was still focused on a domestic chatbot, ChatGLM, and she owned the strategy for that product.

Zhipu, also branded Z.ai, was founded in 2019. The AGI bet at the time had nothing to do with large language models. The first product, AMiner, used graph networks to stitch journals, papers, scholars, and citations into a searchable map of academic contribution, a kind of Google Scholar with a graph underneath. It was popular in research circles. In 2020 the company shifted to language models, and in 2021 it published the GLM paper — a year ahead of GPT-3.5, Li says, and a new architecture rather than a mild fork of Transformer.

International developers noticed later. In 2024 the whole industry stared at Chatbot Arena; Z.ai ranked somewhere between sixth and ninth by chasing which of two answers a user preferred. In 2025, after Anthropic shipped Claude Code, the team decided coding and agents created more economic value than free chat. Chat on Z.ai remains free. People pay for coding and agent use. GLM 4.5 and 4.6 became known through Claude Code, Kilo Code, and similar tools, not through the chatbot.

The GLM 4.5 technical report describes the method: train three separate teacher models, then distill them into one unified reasoning, agent, and coding model. Pre-training and post-training sit next to each other and share that single target. Li thinks the closeness is why they built 4.5 more efficiently than peers who keep those teams further apart.

Western developers often read Chinese open-weight releases as ideology. Li's account is more practical. Llama, Qwen, and Kimi all contribute research, she says, and Z.ai wants that too. Beyond that, a Chinese company has to be open to be accepted. People will not try the model through the company's API. They will deploy it on Fireworks, run it on Groq, or download it onto their own chips. Without weights, GLM never enters the conversation.

Lambert, who works at Ai2, agrees from the other side of the table. He cannot sign up for a Chinese enterprise API, but he has distilled several Chinese models while training. Li adds the plot twist: Z.ai's own 2024 flagship was closed. DeepSeek R1 showed them you can get famous by open-sourcing and still collect API revenue and partnerships. "You need to expand the cake first and then take a bite."

Making money after the weights are free takes two layers. One is subscription. The GLM Coding Plan charges a monthly fee so users stop watching every prompt burn tokens; a single Claude Code round, she says, can consume a million tokens. The other is engineering on top of the same checkpoint: faster decoding, search, a fuller MCP toolchain. Customers ask why they should not just self-host the open build. The answer has to be something the weights alone do not give them.

She is optimistic about eating into the U.S. coding market without winning it. She does not need to persuade 50% of Claude Code users. Five percent is already a huge market, and the market itself is growing fast. Coding is not the only bet. Role-play is another, and she jokes that maybe one day even Manus will be running GLM.

Asked what distinguishes the lab, Li starts with focus. Everyone works toward one target. Pre-training and post-training have separate heads, but they sit together. The core staff actually training the model is 100 to 200 people, the same band U.S. labs talk about. That is enough, she says, because the core has to stay concentrated and cannot afford many mistakes. Data prep and product sit around them.

Bigger Chinese companies look different only at the portfolio level. ByteDance can chase text, video, and speech at once and spread GPUs across teams. Inside each team the pattern repeats: 10 to 20 core people, another 80 to 100 on training and data. Large firms pick talent first because they can pay more. Startups, Li says, need people who will fight together for less money and treat building GLM as the point.

She keeps returning to one management rule. The head of the team has to do the research and the training, not just set goals. Things change too fast. Mid-run, Claude 4 or GPT-5 can land, or a rival can ship a new architecture. You have to feel the move yourself. Even the founder reads papers and runs experiments. Many people on the team are current PhD students who treat GLM as their biggest academic result, not as a side job to finish after the dissertation.

Schneider asks about the media story that DeepSeek was biased against people who studied abroad. Li sees no such bias at Z.ai. Interviewers want the best performance in the room; domestic candidates often win because they interview better on that rubric, not because the bar is different. She joined after MIT, and almost nobody knows. People never ask whether you have a foreign master's. Maybe ten colleagues have heard she studied there. "There's no bias because people don't care."

Lambert asks how long it takes from a finished training run to a public release. Li's answer is several hours. They run evaluations, then open-source. There is no queue to send an endpoint to LMArena or Artificial Analysis first, and no Nano Banana-style teaser campaign. If you have decided to open the model, she says, the open-source drop itself is the biggest event.

That cadence is hard on the person who has to do partnerships. Li would like a week to work with inference providers, benchmark shops, and coding-agent teams so outsiders can try the model first. The company view is narrower: prepare the open-source materials, the benchmarks, and a tech blog, then ship. She often gets a late-night ping that a new model is two or three hours away and has to wake Fireworks, LMArena, and Kilo Code inside that window.

Hours inside the company vary. Some people leave at 7 p.m.; some never leave. Li works about 18 hours because she has to meet U.S. CEOs and coding-agent founders at 2 or 3 a.m. Researchers and engineers are a different case. Brains, she says, can do maybe eight hours of papers, code, and experiments. Anyone who can sit in meetings for 20 hours is attending meetings, not training a model. Lambert's old PhD advisor used to say four hours of real technical work can change the world if you walk in the sun after that.

The company metaphor is a long run. The founder is a serious marathoner, under three hours. For the anniversary they staged a half marathon. An intern — now the product manager of Z.ai chat — finished in 3 hours 15 minutes. Li waited at the line; the intern, she says, was almost dead, then ate two hamburgers. Energy matters when the hours are long.

GLM 4.5 was relatively weak at role-play because the team had not trained on that data. The genre uses very long system prompts. Users will write five pages of backstory — Schneider's example is a Japanese pirate raiding the Taiwan coast in 1570 — and then play a text RPG. If the model was not trained on that length of instruction, it forgets who it is and falls back to generic chat. Li says the fix is mostly post-training data, and 4.6 has a lot more of it.

One internal experiment lets a user describe what Stewie does on Family Guy and generate a personal Stewie. Text already works; a speech model would recreate the voice. Asked about AI psychosis and people falling in love with chatbots, Li has read OpenAI's recent work on making models less addictive, including training them to say they are AI. Internally people discuss it when it hits the news. For a broader Chinese audience she thinks the topic is early. Performance is still not on par with the top closed models, and behavior changes so much between versions that safety work on last week's checkpoint goes stale.

Translation is the other domestic specialty. On Chinese-English, Li puts GLM on par with Gemini 2.5 Pro. The edge is memes and emoji. If a sentence is about AI and uses a whale to stand in for DeepSeek, the model should render "DeepSeek." If the sentence is about animals, it should stay a whale. Chinese internet speech is full of abbreviations, animal brand names, and naughty comment-section slang. WeChat and Xiaohongshu data sit behind walls, so the team uses public observation plus synthetic data. TikTok comments, she says, are where people are most mischievous, and the TikTok-refugee wave even brought enterprise translation customers.

Accuracy requirements split by product. YouTube auto-translate can be 80% right; users only need the gist. Comments on X, Xiaohongshu, and WeChat have to be essentially 100% or the customer will not buy. The same cultural work is harder beyond Chinese and English. The models already cover about 20 languages, including French and Hindi, but after that the team has to register for local apps just to see how people talk.

Lambert asks the question American trainers usually dodge: how much further can today's transformers go if you just improve the data, and where is the wall? Li does not dodge. Data can improve results, but it cannot cross the wall. There is a wall. They need better architecture for both training and post-training. GLM 4.6 is a 355-billion-parameter model, too expensive to use as a sandbox, so hypotheses get tested on 9-billion or 30-billion runs. Ninety percent of those experiments fail. That, she says, is ordinary scientific work.

Another industry fiction is the context window. A model can advertise a million tokens and still be reliable only inside 60,000 or 100,000. The product fix is unromantic: compress the window to 30,000 or 60,000 so the user does not feel the drop. People do not usually need a million tokens anyway. Single-agent setups still win on Z.ai chat for now, because one strong model can see the whole thread. Multi-agent orchestration means compressing context per agent; one hallucination can ruin the chain.

All training stays in China. Inference for overseas customers runs in Singapore on api.z.ai, operated by a Singaporean company — Li is technically an employee of that entity — and data does not bounce back to the mainland. The reason is not a regulatory dodge. Overseas customers demand local storage, and the privacy policy gets stricter every month. They are not routing training through AWS or Nebius because they already have chips at home and a simpler stack than OpenAI or Anthropic. Blackwell would be welcome for FP4 cost cuts; Huawei and Nvidia chips get sorted by use case, 30 tokens per second versus 80.

Zhang asks how Chinese society talks about AI risk. Li's observation is specific: developers feel the fear first, especially juniors. They live in Claude Code and Codex and can see how much of the job those tools already do. Writers and managers already had SaaS and junior staff. A chatbot is another helper. Ordinary DeepSeek users brainstorm, polish prose, and translate, and they do not believe the tool replaces them.

The fears she hears are about jobs, not about AI taking over the world. Lambert notes the contrast with the United States, where a vocal minority treats job loss as a given and then piles on deeper worries about power and control. Li lived through that ecosystem. At MIT, she says, everyone talked about how AI would change the world, and not on the positive side. Zhang asks why China sounds more practical. Li's answer is coverage. Maybe a million people follow the latest trend. More than a billion go to work and are not yet in the loop. "The more you learn, the more fear you will have."

Inside Z.ai there is no fear of unemployment, because the work exceeds the headcount. Large enterprises are a different story. A company with 10,000 people in data analysis and backend engineering might need half of them once the boss picks an agent stack. Ordinary employees cannot do much about that. They wait for the decision, the way people wait through Amazon layoffs. Until models stop hallucinating and start following instructions, Li thinks most of the public will keep calling the technology a bit silly rather than terrifying.

A German professor talking to Li the day before the taping named Llama, Qwen, even Mistral — and not GLM. Lambert counters that in San Francisco circles, more people now talk about GLM than about Mistral, and arguably more than Llama. Progress, Li agrees, but Reddit still fills with "what is GLM?" posts. Z.ai has about 20,000 followers on X. DeepSeek has a million, a number Lambert says would be large even for a new American tech company. Mistral and Cohere still get more attention than Kimi or Z.ai.

Chinese tech media watch Karpathy, Sam Altman, and Elon Musk so closely that a single tweet about Claude Code versus Codex ricochets home the same day. Li says Z.ai used to neglect the global brand because the plan was to sell APIs to Chinese enterprises. Those enterprises now watch overseas reputation before they buy. Global mindshare is no longer optional marketing. It feeds domestic deals.

Enterprise sales inside China split in two. Some firms will not send data to Z.ai or even Alibaba; they want DeepSeek deployed on their own chips, usually with RAG and workflows bolted on by a third party. Others — tech and media companies that need standardized workflows — will take an API and pick on price versus performance. ByteDance, she says, dominates that API market. Qwen 3 Max stays closed so Alibaba has something to sell. Z.ai open-sources its flagship, which is why customers keep asking what the paid service adds.

Overseas demand is not only American. Users come from India, Indonesia, Norway, and Brazil, mostly via X, Reddit, and YouTube. India has the most users. The United States still produces about half of overseas revenue, because Americans buy Pro and Max plans while Indian users buy Light. Zhang asks how Li explains the work to a Beijing high-schooler or a grandmother. It is hard. She can only say they do something like DeepSeek. Even many Tsinghua students have not tried GLM, or have not heard of the company. "Only the most famous one get all the attention."

Li will name the research bets even if she will not name GPU counts. The team is mature on off-policy reinforcement learning and is now exploring on-policy RL. Multi-agent architecture is the other live file. Z.ai chat already searches, then searches again, then builds slides or a poster, but one GLM 4.6 actor does all of it. A solid eval harness matters because they have several product shapes and need to know whether more agents actually improve speed and quality.

The near-term lineup is smaller, not larger. GLM 4.6 Air and a Mini around 30 billion parameters were due within weeks of the taping, plus a 4.6 vision model. That is the rest of 2025. 2026 work is still on small models and will not become a product in that year; it is there to teach them how to train the next generation. Search on the Chinese internet remains a builder problem, she adds, the same way Google and Bing starve Western products of clean APIs. Scrapers and browser agents in the Manus style are more realistic than hoping for an official feed.

There is no grand U.S.-China rivalry speech at the end. Li's self-portrait is a lab that is not OpenAI or Anthropic, still in an early stage, using the chips it has. If Z.ai can hold a meaningful niche in a global market, that is already enough. Open weights exist so the model gets used. Releases move in hours because resources are thin. People are not terrified yet because the models still hallucinate. Those constraints, more than any slogan, are the picture of a Chinese frontier lab in late 2025.