Tang Jie personally spoiled GLM-5

Tang Jie personally spoiled GLM-5.5: The phrase "epic plus" has raised the arms race of domestic large-scale models to a higher level. The intensive outbreak of domestic large-scale models is bringing the position of the US flagship closed-source AI so close that it is difficult to distinguish it with the naked eye. Kuai Technology reported on July 20 that after Qianwen and Kimi showed off their muscles one after another, it was Zhipu’s turn to take action. The rumored GLM-5.5 was described as an epic upgrade. But this time it was not media speculation, but Tang Jie, the founder of Zhipu himself. Some netizens mentioned in the comments that Qianwen and Kimi have both completed epic evolutions, and casually asked if GLM still has a future. Tang Jie's response immediately raised expectations - he used one word: epic plus. This means that in Zhipu’s own judgment, the improvement of this generation is greater than those mentioned by netizens. Of course, it is impossible to talk about specific figures now, but Tang Jie obviously hopes that the outside world will maintain confidence in GLM. His confidence comes from the record of the previous generation: before Kimi K3 debuted, GLM-5.2 was the first player among domestic large models to equal Opus-level performance in terms of AI programming capabilities. What is even more noteworthy is its size - the parameter scale of GLM-5.2 is only more than 740 billion, which is almost a quarter of Kimi K3, and probably only one-fifth or less of the top AI in the United States. However, it has relied on backward training to push its capabilities to that level. It can be called a sample of small parameters leveraging large capabilities. The new epic plus model mentioned by Tang Jie should point to the next generation GLM-5.5. The name GLM-5.3 has been rumored in the circle before, but now it is more likely to be GLM-5.5. Of course, it is not ruled out that it will be directly named GLM-6 when it is released. A certain trend is that the number of parameters of the next generation of large models will at least cross the 1 trillion threshold. Considering that DeepSeek open source base model has been used in Zhipu’s history, it is speculated that GLM-5.5 may achieve the same level of 1.6 trillion parameters as V4Pro; coupled with Zhipu’s recognized strong post-training skills, its performance jump is indeed worth betting on, and it is likely to be another large domestic model that can compete with the Fable5 level. However, in the waiting list of the domestic camp, the most uneasy thing right now is the official version of DeepSeek V4. When the report was issued, the last moment of mid-July - zero o'clock that day - was hanging overhead. No one could tell whether the little whale planned to make a big move in the middle of the night or jumped again. When Zhipu ignited the topic with an epic plus, the relay race of domestic large models was obviously far from reaching the finish line. via AI News (author: AI Base)