Investment community summary says Zhipu's next-generation model may reach 3500000000000
BIBIBI
AT A GLANCE
Circulating summary says Zhipu's new foundation model may reach 2500000000000 to 3500000000000 parameters, but the parameters, progress, and capability projections have not been confirmed.
Article
Zhipu's new model reportedly reaches up to 3500000000000 parameters, potentially reaching the Claude Fable tier
Beating AI Flash: An expert interview summary circulating in the investment community states that Zhipu's next-generation foundation model is not the previously rumored 1600000000000 parameters, but rather reaches 2500000000000 to 3500000000000 parameters.
The August released by Zhipu this year GLM-5.3 still uses the previous-generation foundation model, with approximately 744000000000 total parameters. If the information is accurate, the new foundation model will be approximately 3.4 to 4.7 times the size of the current version.
However, the training process has not been smooth. The summary states that Zhipu adopted a more aggressive model architecture this time, and during pre-training, there was at one point a situation where Loss did not decline—meaning compute continued to be invested while training metrics could no longer improve.
The interviewee believes the issue should have been resolved recently and expects training to be completed soon, predicting that the new model can reach flagship-level performance comparable to Claude Fable.
The new model also aims to address the issue of "being able to launch but unable to meet demand" after large-model releases. The summary states that Zhipu treated inference speed and cost as hard constraints from the architecture-design stage, with the goal of providing services at scale on the first day of release, and that it has already discussed revenue sharing with multiple cloud providers.
Original link https://m.theblockbeats.info/flash/372939
Key points
01
An expert interview summary circulating in the investment community states that Zhipu's next-generation foundation model is not the previously rumored 1600000000000 parameters, but rather reaches 2500000000000 to 3500000000000 parameters.
02
The original article states that the August released by Zhipu this year GLM-5.3 still uses the previous-generation foundation model, with approximately 744000000000。
03
Based on this, the original article calculates that, if the information in the summary is accurate, the new foundation model would be approximately 3.4 to 4.7 times the size of the current version.
04
The summary states that during pre-training of the new model, there was at one point a situation where Loss did not decline, meaning that compute continued to be invested but training metrics could no longer improve.
05
The interviewee believes the issue should have been resolved recently, expects training to be completed soon, and predicts that the new model can reach the Claude Fable tier.
06
The summary states that Zhipu treated inference speed and cost as hard constraints from the architecture-design stage, with the goal of providing services at scale on the first day of release.
07
The summary also states that multiple cloud providers have already discussed revenue sharing.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This information comes from an interview summary circulating in the investment community, rather than the official statement presented in the article. It claims that Zhipu is training a larger next-generation foundation model and attempting to balance inference speed, cost, and service capacity; however, the original article provides no verifiable results regarding whether training has been completed, the actual parameter count, or final capabilities.
Why it matters to readers
If the targets in the summary can be achieved, the new model may see changes in scale, inference availability, and service costs; however, only secondhand accounts and the interviewee's projections are currently available, so model performance or commercial partnerships cannot be confirmed on this basis.
Beginners should distinguish between "statements in the summary," "the interviewee's assessment," and official facts that have already been released. More parameters do not necessarily mean stronger capabilities, and whether it reaches the claimed tier still requires verification through post-release technical materials and real-world evaluations.
Risks and unknowns
Whether Zhipu has confirmed 2500000000000 to 3500000000000 parameters remains unclear.
Loss It has not been confirmed whether the issue of failing to decline has been resolved.
The training completion date is only an estimate; actual progress is unknown.
Reaching the Claude Fable tier is the interviewee's prediction and lacks evaluation results.
Related Developments
Loading event timeline…
Whether the goals for inference speed, cost, and large-scale service on the first day can be achieved is unknown.
The identities of multiple cloud service providers, the progress of negotiations, and revenue-sharing arrangements have not been disclosed.
The original text does not provide official announcements, technical reports, or partnership announcements from Zhipu.
Related concepts
Foundation model
This term is not in the glossary yet. Browse related concepts in the glossary.
