Rumored meeting notes put Zhipu's next base model at 2.5T to 3.5T parameters on an aggressive architecture, with a “Fable-level” capability target and training said to be close to done. Zhipu has confirmed none of it. The verified record around the rumor: a GLM HighSpeed trial in ZCode 3.15.1 (Oct 12), a “glm-5.5-vision” string in a user screenshot (Oct 10), the GLM-Omni paper (Oct 15), and an RSI engineering post describing how GLM-5.3 built its own serving stack (Sep 17).
- Rumored scale: 2.5T to 3.5T parameters, unverified
- Stated target: “Fable-level” flagship capability
- Trial sighting: GLM-5.3-Highspeed at ~490 tokens/s
- Zhipu's confirmation: None as of Oct 22, 2026
October has not been kind to the uptime logs of China's model providers. In the early hours of October 22, some GLM App users could not chat at all; a support reply shared by users puts the outage at 00:56 to 04:14, about three hours and eighteen minutes, with service restored afterward. Days earlier, DeepSeek had worked through its own disruption and returned on October 17 with an upgraded DeepSeek-V4.1-Flash-1017. Around the same stretch, a set of rumored meeting notes about Zhipu's next base model started circulating. This piece keeps what those notes claim separate from what Zhipu has actually put on record.
How big is Zhipu's next base model rumored to be?
The notes put the next base model at 2.5 trillion to 3.5 trillion parameters. They describe the architecture as aggressive and sized to serve large-scale traffic on the first day of release. Training reportedly hit a stretch where the loss refused to come down; the notes say that problem has since been resolved and training should finish soon. The capability target is described as “Fable-level” flagship quality.
None of that is confirmed. Every number in this section comes from a leaked document, and Zhipu has published nothing that would back any of it.
Which rumors does the leak itself push back on?
The notes were also asked about an earlier rumor: GLM-5.5 launching in late October with 1.6 trillion parameters. The reported answer opens by calling that claim inaccurate. So the same document that inflates the parameter count also cuts down the most specific date attached to the GLM-5.5 name.
One more detail from the Q&A is worth keeping: GLM-5.3 shares its base model with GLM-5.2, and its gains came from post-training. If the current flagship's strength comes from post-training rather than a new base, a 2.5T to 3.5T base would be a much larger step than the last one. That is exactly why it needs independent confirmation.
Every GLM-5.5-adjacent claim and its status as of October 22, 2026
| Claim | Status | Where it surfaced |
|---|---|---|
| Next base model at 2.5T-3.5T parameters | Unverified rumor | Leaked meeting notes, Oct 22 |
| “Fable-level” capability target | Unverified rumor | Same notes |
| GLM-5.5 in late October at 1.6T | Disputed in the leak itself | Earlier social media rumor |
| GLM HighSpeed trial for subscribers | Official | ZCode 3.15.1 changelog, Oct 12 |
| “glm-5.5-vision” string | User screenshot | Community forum post, Oct 10 |
| GLM-Omni multimodal model | Official | Tsinghua & Zhipu paper, Oct 15 |
What is GLM HighSpeed, and how fast is it?
On October 12, ZCode shipped version 3.15.1. The changelog says subscribers may be randomly selected to take part in a GLM HighSpeed trial. Users drawn into the grey launch report the model string GLM-5.3-Highspeed and output speeds around 490 tokens per second.
That figure is one user's measurement from one trial wave. Zhipu has published no benchmark, and speeds in grey-launch trials vary with prompt length and load. The timing is the interesting part: a serving-speed experiment weeks before a rumored flagship launch looks like preparation for day-one capacity, though that reading is ours rather than anything Zhipu has said.
Why did a “glm-5.5-vision” string surface in October?
On October 10, a user on the LINUX DO forum reported that while using GLM-5.3 through a third-party client called Pi, an assertion failure popped up carrying the string “glm-5.5-vision”. The evidence is a screenshot of an error message. Internal strings get tested, reused, and dropped all the time, and a name inside a failed assert is not a product plan.
Still, a vision-prefixed GLM-5.5 would line up with the company's multimodal push that same month, because five days later Zhipu's own researchers put a multimodal model on paper.
What is GLM-Omni?
On October 15, Tsinghua University and Zhipu published a paper titled “Can Visual Generation Expand What Unified Multimodal Models Know?” It discloses GLM-Omni, an internal unified multimodal model, with Tang Jie as corresponding author. The research question is in the title: whether generating visual data can expand what a unified multimodal model knows.
GLM-Omni is a research model, not a product announcement. But it is the clearest official signal of where Zhipu's multimodal stack is heading, and it gives the “glm-5.5-vision” screenshot a plausible home.
How far has Zhipu's recursive self-improvement work actually gone?
Only one part of this story has a published trail behind it. On September 17, Zhipu's engineering blog described an Infra Agent driven by GLM-5.3: starting from a first run on new hardware, it pushed GLM-5.3-Flash into production serving in about two weeks, lifted end-to-end throughput by roughly 3x, and did so across a fleet of more than 100,000 domestic accelerators.
Tang Jie's October talk in Singapore added the direction: pre-training, post-training, and agents all keep scaling. Recursive self-improvement has produced results (the Infra Agent is one) but is not yet a mainline production method, and Loop-related techniques may reach the serving side first.
The link to the rumor is direct. If a 2.5T to 3.5T model is real and meant to serve heavy traffic on day one, the hard part is the serving layer, and that is the layer Zhipu has been publicly rebuilding with its own models.
What comes after the flagship?
At Zhipu's H1 review on August 31, Tang Jie described the GLM-6.0 direction as Full Self-Training. Whatever GLM-5.5 turns out to be, whether a 3.5-trillion-parameter flagship, a vision variant, or nothing at all, the company's stated trajectory runs toward models that train more of themselves.
This article summarizes an unverified leak and community reports. Zhipu has not confirmed the model name, the parameter count, or any release timing. The numbers above may change or turn out to be wrong.
Our coding plan
Rumors are fun to follow, but they don't write code. If you need a coding model that ships today, our Coding Plan comes with official API keys, a direct connection, and a deep discount off list price, while stock lasts.
Frequently asked questions
Has Zhipu confirmed GLM-5.5?
No. As of October 22, 2026, Zhipu has not confirmed the model name, the parameter count, or a release date. Every number sizing the next flagship comes from unverified leaks or community reports.
How large is the rumored next base model?
Between 2.5 and 3.5 trillion parameters, according to meeting notes that circulated on October 22, 2026. The same notes dismiss an earlier rumor of 1.6 trillion parameters with a late-October launch.
What is GLM HighSpeed?
A speed trial opened by the ZCode 3.15.1 changelog on October 12, 2026. Randomly selected subscribers see the model string GLM-5.3-Highspeed, and one user report puts output around 490 tokens per second.
What is GLM-Omni?
A unified multimodal research model disclosed in an October 15, 2026 paper from Tsinghua University and Zhipu, titled “Can Visual Generation Expand What Unified Multimodal Models Know?”, with Tang Jie as corresponding author.
Is GLM-5.3-Highspeed the same thing as GLM-5.5?
There is no evidence that it is. The trial string says GLM-5.3, and Zhipu has not connected the HighSpeed trial to any future model.