Listening to the rain, from Aofeisi Temple
QuantumBit | Public Account QbitAI
Zepure has introduced a new GLM-5.3-Highspeed model, which is available in Zcode.
According to netizen @leo114119, this new model is currently available for trial in Zcode 3.15.1, and subscription users can get it randomly.
The speed it generates is even up to 490 tokens/s.
I also tried it in Zcode right away, and attempted to manually add and switch this model.
But I don’t know whether it’s because users on the Lite plan don’t meet the eligibility criteria, or because they haven’t fully adopted the grayscale feature 😭...
Only able to sigh in despair first(:з」∠)
GLM-5.3 High-Speed Version Is Here?
Since it’s a high-speed model, what’s most important to focus on is, of course, speed.
490TPS—The speed figure posted by this netizen is also quite terrifying, close to 500TPS.
TPS refers to the number of tokens generated per second. If this rate is maintained, producing 10,000 tokens would take approximately 20 seconds for the generation phase alone.
You should know that the GLM-5.3-FlashX, just released by Zhipu last month, has a maximum speed of only 200 tokens/s.
And according to the speed measurement by Artificial Analysis, the GLM-5.3-Flash outputs at approximately 60 tokens/s on the Z.ai API.
What speed does this new model have, immortal guy???
And the name GLM-5.3-Highspeed has shown signs before.
The Pi model directory has included GLM-5.3-Highspeed.
In this directory, the GLM-5.3-Highspeed context window is 1 million tokens, with a maximum generation token limit of 131,072, approximately 128k; it also has three levels of thinking intensity: low, high, and max.
However, these configurations are also consistent with the GLM-5.3 standard version.
In addition, some netizens have discussed this new model on Reddit, and speculated whether it is related to the anonymous model Omen Alpha on OpenCode.
Omen Alpha just released OpenCode in September, and it has been revealed by netizens to be related to GLM.
Some netizens also tested the pelican’s cycling skills..
In summary, it is: fast, takes a long time to think about, but the effect is quite good.
As for what exactly is the Omen Alpha model, there are also various opinions.
Some people say it is GLM-5.3-Highspeed, while others guess GLM-5.3-Omni or GLM-5.4-Flash.
Previously, Zhipu also launched a high-speed version of the model.
In May this year, Zhipu launched the GLM-5.1-highspeed, with a model output speed of 400 tokens/s. At that time, it also basically exceeded the upper limit of the API speed for global large model manufacturers.
If the GLM-5.3-Highspeed can actually reach 490 TPS this time, then the speed would be quite impressive for long-distance tasks.
Currently, this new model is still in random grayscale mode in Zcode. If any netizens have seen it in grayscale, feel free to share your actual usage experience!!