Zhipu's Year-End Run for the Top: Are 3 Trillion Parameters Enough?
Word from Zhipu's shareholder meeting: the 700-billion-parameter GLM 5.4 Preview is now in testing, and the 3-trillion-parameter GLM 5.5 will ship at year's end. That's a 4x jump in parameters — and plenty of skepticism to go with it.
There's now a clear account of Zhipu AI's release cadence for its new models.
According to someone who attended Zhipu's November 13 shareholder meeting, combined with a JPMorgan meeting note that leaked recently, the roadmap for the new GLM models looks roughly like this:
First, a 700-billion-parameter GLM 5.4 Preview tests the waters. That model is already being tested in a limited circle.
The headline act is GLM 5.5 in December, with a target parameter scale of 3 trillion for the core model — roughly 4x the 745 billion from January of this year — aiming to break into the global top tier.
Source: SuSu_酥酥's post
Source: SuSu_酥酥's post
Zhipu AI is one of China's domestic large-model companies, and the GLM series is its core product line. Over the past few months, rumors about GLM-5.5 have never let up. A while back some said it would go live on the 19th–20th, but nothing materialized. This time, the shareholder meeting plus the JPMorgan note adds up to the most concrete roadmap so far.
But once the news spread, the comments section split into camps.
One camp questions the parameter race. One commenter said that most of the progress in commercial models over the last six months cannot possibly be due to larger parameter counts — that's going down the wrong path. Others cautioned that 700 billion parameters sounds impressive, but inference cost, stability, and real-task performance are what actually determine whether something is usable day to day. Quadrupling parameters also means real hardware pressure. One user did the math: the GPUs will start smoking first, and running a demo will probably have to wait until after the New Year.
Source: dragon's post
Another camp questions the news itself. One commenter pointed out that discussing this much detail at a shareholder meeting is a compliance taboo: a Hong Kong-listed company hasn't issued an announcement, and briefing only some shareholders on the roadmap is most likely someone filling in the blanks after the meeting. Some even said that, reportedly, the shareholder meeting didn't discuss any of this at all.
This kind of news has already made the rounds in the Chinese AI community, with different accounts reposting each other, and it's hard to tell what's true. Secondhand accounts plus a leaked note can piece together a roadmap, but they can't piece together the truth.
The more sobering lesson is in the stock market. One commenter said they believed the Zhipu news, bought into GLM's Hong Kong shares, and lost 90%. Model strength and share price have never been the same thing. Even if GLM 5.5 turns out exactly as rumored, for those already underwater it's just emotional comfort.
Zhipu's playbook itself isn't wrong: test the waters with 5.4 first, then go for the top with 5.5 at year's end. But whether it can hold a place in the top tier — parameters are only the entry ticket. Post-training, inference cost, ecosystem support: if any one link in the chain fails, 3 trillion won't save it.
See you in December.
发布时间: 2026-11-22 21:42