Citation Bureau
Vol. I
No. 389
XVII SEPTEMBER MMXXVI
Reference

What is ZAI?

ZAI

ZAI is a Chinese AI lab that develops the open-source GLM series of large language models. The material tracks the lab’s GLM releases and its stated ambitions around recursive self-improvement and open-source model capability.

How it developed

  • Jun 2026 - Nathaniel Whittemore said GLM 5.2 ranked first on Bridgebench and first on reasoning, beating Fable 5 at one-tenth the cost and 300 tokens per second.
  • Jul 2026 - ZAI CEO Zhi Tang said the lab was not there yet, though it expected to have an open-source model at Mythos level by the end of the year.
  • Sep 2026 - Nathan Labenz said it was the first time a Chinese lab had openly declared it would build recursive self-improvement.

In the evidence

Every line below is attributed to a named speaker.

Company & tool watch

ZAI, a Chinese lab, is the first in China to openly declare it will build recursive self-improvement, marking a notable public escalation in AGI ambitions from a Chinese actor.

“This is the first time a Chinese lab has openly declared they will build recursive self-improvement.”
Nathan Labenz · 16 Sep 2026
By the numbers

GLM 5.2 from ZAI ranks number one on Bridgebench and reasoning, beating Fable 5 at 1/10 the cost and 300 tokens per second throughput.

“GLM 5.2 is number one on Bridgebench and number one on reasoning beating Fable 5 at 1/10 of the cost and 300 tokens per second.”
Nathaniel Whittemore · 19 Jun 2026
Company & tool watch

GLM 5.2 from ZAI tops Bridgebench and reasoning benchmarks while undercutting Fable 5 on price by 10x, positioning ZAI as a serious frontier challenger.

“GLM 5.2 is number one on Bridgebench and number one on reasoning beating Fable 5 at 1/10 of the cost and 300 tokens per second.”
Nathaniel Whittemore · 19 Jun 2026
Company & tool watch

ZAI, whose CEO Zhi Tang expects the company to release an open-source model at Mythos level by end of 2026, is a frontier open-source lab to watch.

“Even the ZAI CEO Zhi Tang said that they are not there yet, although they do expect to have an open-source model at Mythos level by the end of the year.”
Zhi Tang · 14 Jul 2026
Citation Bureau · reference note, compiled from attributed expert discussion. Last updated 2026-09-17.