China's AI Upstarts: How Z.ai Builds, Benchmarks & Ships in Hours, from ChinaTalk

**SPEAKER_1** (0:00)
This podcast is sponsored by Google. Hey folks, I'm Amar, Product and Design Lead at Google DeepMind. We just launched a revamped vibe coding experience in AI Studio that lets you mix and match AI capabilities to turn your ideas into reality faster than ever. Just describe your app and Gemini will automatically wire up the right models and APIs for you. And if you need a spark, hit I'm feeling lucky and we'll help you get started. Head to ai.studio slash build to create your first app.

**Nathan Labenz** (0:31)
Hello, and welcome back to The Cognitive Revolution. Today, I'm honored to share a special cross post from the ChinaTalk podcast, hosted by Jordan Schneider, ChinaTalk analyst, Irene Zhang, and Nathan Lambert of AI2 and the Interconnects sub stack. Featuring a conversation with Zixuan Li, Director of Product and Gen. AI Strategy at ZAI, also known as Zhipu AI, about the culture, incentives, and constraints shaping Chinese AI development. Now, I imagine that many, even in our AI-obsessed audience, will not be familiar with ZAI, but their model releases demonstrate that they are a significant player worthy of our attention. As of today, their latest GLM 4.6 model holds the number 19 spot on the LM Arena text leaderboard. Its ELO rating is roughly 65 points behind the current leaders, which means that it still wins 2 out of 5 head-to-head comparisons with the leaders, and it happens to sit right next to Quen 3 Max, Kimi K2 Thinking, and DeepSeek's V32. Together, these four models, all from China, are indeed the top four open-source models available today, though it should be noted that Mistral is not too far behind. On the web development leaderboard, GLM 4.6 does even better, coming in at number 9, making it competitive with GPT 5.1 and meaningfully behind only Gemini 3 and the new Claude 4.5 Opus. All that said, this conversation goes way beyond benchmarks and touches on a number of important topics, including why we should understand Chinese companies' open-weight strategy not so much as an ideological commitment, but as a practical marketing tactic from companies that are seeking to gain global mind-share while recognizing that Western enterprises simply can't and won't use their APIs. The culturally distinct AI use cases such as role-play that matter in China, and how these drive different fine-tuning priorities than what we typically see from Western companies. The role that Silicon Valley thought leaders play in establishing credibility for Chinese companies, even in their home market. The market for AI talent in China, and why very few people at ZAI even know that Zixuan studied at MIT. Zixuan's view that there is a wall, and that further architectural breakthroughs will be needed. The extreme velocity with which ZAI releases models, which often involves shipping within hours of completing training. What, if anything, people in China fear about the AI future, and how Chinese companies generally still view themselves not as peers or rivals with leading American companies, but as upstarts that are just trying to keep pace, and who will be quite happy if they can secure for themselves a meaningful niche in the global AI marketplace? This is a fascinating, detailed, rich conversation, and regardless of your attitude on US-China competition, it's clear to me that we in the West don't hear nearly enough from the actual builders inside Chinese AI labs. So I appreciate Jordan for allowing me to cross post this episode, and of course I encourage everyone to subscribe to ChinaTalk, online at Chinatalk.media.
With that, I hope you find as much value as I did in this behind-the-scenes look at Frontier AI Development in China. With Zixuan Li of ZAI from the Chinatalk Podcast.

**Jordan Schneider** (3:44)
Zixuan Li, who studied in the US before moving back to China, works at Drupu AI or ZAI. We're going to let him introduce himself in his role, co-hosting today Irene Zhang, a longtime Chinatalk analyst, as well as Nathan Lambert of AI2 and the Interconnects sub-stack. Welcome to Chinatalk, everyone.

**Nathan Lambert** (4:07)
Personally, I've known of Zhipu AI or ZAI for at least over a year, and I've been following the work there. Then December, it was like, kind of in my mind, it was like, well, there's another deep siege when they released GLM 4.5. You'll have to correct me on whether or not it was before or after the Kimi K2 model. My guess is after. That was like the sequence of order of weeks or days after Kimi K2, and it's just like, wow, like, okay, there's a lot of people building great models, and it's fun to get to learn about some of them, and how it compares between US labs and Chinese labs, and I think a lot of it is more in common than different.