Domestic AI performance benchmark Kimi K3. released in mid-January: K4 aims to reach 5-6 trillion parameters

Kuai Technology reported on October 11 that as the current strongest benchmark for domestic AI, Dark Side of the Moon’s Kimi K3 has been released for almost three months. The upgraded version K3.1 was rumored at the end of last month and is currently expected to be released in mid-October. According to information from Nomura Securities, Kimi K3.1 is planned to be released before mid-October, focusing on improving the Agent's ability to continuously perform long-term tasks. The team plans to further optimize the model using post-training, computer operation trajectories and real mission data. The current Kimi series of large models are mainly divided into two directions. Flagship models such as K3/K3.1 continue to iterate on the intelligent upper limit. Kimi K2.8, ​​released on September 11, focuses on high efficiency, focusing on writing code and daily load of agents. In addition to K3, Dark Side of the Moon is also training a more powerful large model with an estimated parameter volume of 5-6 trillion parameters, which should be the previously rumored Kimi K4. For the Kimi large model, the current main problem is insufficient computing power, especially the lack of high-end computing power. Domestic AI chips are not a big problem for daily reasoning, but large-scale training cannot replace the NVIDIA graphics card in the short term. We previously reported that after the National Day holiday, a domestic AI outbreak is expected. Now the first half of October has passed. If it is soon, we will see the first new products released next week. Kimi K3.1 personally believes that it is the most anticipated one after DeepSeek V4.1 Pro. Because the previous Kimi K3 has already been the performance benchmark of domestic AI, with 2.8 trillion parameters, it is also the largest in domestic and even open source fields. When it was released, it was also a large domestic model whose performance truly caught up to the Fable level. Especially after the official version of DeepSeek V4 Pro was not satisfactory, its significance became even more important, and it also carried the banner of the best performance of domestic AI. However, Kimi K3 has had many problems in the past three months. The first is usability. It once reached the point of limited-sale packages. Recently, it has improved slightly, but it is not really satisfactory. Secondly, the price of K3 is also