Kimi K4 is expected to be larger than K3, but the dark side of the moon is being blocked by an NVIDIA chip.
According to DeepInsight Beating monitoring, the dark side of the moon has already started discussing Kimi K4, with the new model's scale expected to significantly exceed that of K3. However, the release date for K4 has not been announced yet, and the company is currently seeking more NVIDIA Blackwell chips to prepare computing power for future development.
《The Information》 reported that part of K3's training was actually completed in China, not all in the Thailand data center as claimed by U.S. officials.
The dark side of the moon has at least accessed Blackwell chips from two Chinese cloud providers. As no single provider can offer sufficient computing power, the team had to connect a large number of eight-card servers and optimize communication between different vendors and data centers.
This stitching solution helped the dark side of the moon complete the K3 with 2.8 trillion parameters, but it also highlighted the challenges for K4. The U.S. has banned the export of Blackwell chips to China and is investigating channels through which companies like the dark side of the moon obtain advanced chips.
The operating computing power after K3 went live is also in short supply. The dark side of the moon heavily relies on NVIDIA H20, with each deployment recommended to connect at least 64 chips. After a surge in user demand, the company suspended new subscriptions in less than 48 hours.
It is currently challenging to temporarily replace imported chips with domestic ones. The supply of large servers that can connect dozens or even hundreds of domestic chips is insufficient, with customers having to wait for up to six months.
K4 currently only has a direction and no timetable. The extent to which it can be developed and when training will begin may depend on how many more Blackwell chips the dark side of the moon can secure.