Moonshot AI (the company behind Moonshot AI) officially open-sourced Kimi K3 on the evening of July 27. The model has 2.8 trillion parameters. At the same time, it released the model weights, a technical report, and key infrastructure. The company also commented almost simultaneously that it is already working on the next-generation K4, which it said will be “clearly larger” than K3, and noted that it urgently needs more Hashrate.
Moonshot AI Open-Sources K3 and Plans for K4: Key Technical Facts
According to Moonshot AI’s public statements and related reporting, the main known technical background for K3 and K4 is as follows:
K3 open-sourcing date: July 27, 2026, with model weights, a technical report, and key infrastructure released at the same time
K3 scale: 2.8 trillion parameters
K3 demand after launch: Within 48 hours of launch, usage reached the limit; Moonshot AI temporarily paused new subscriptions
K3 inference requirements: On H20 chips, at least 64 server chips need to be interconnected
K4 plan: Scale “clearly larger” than K3, actively seeking Nvidia Blackwell Hash Power
Hashrate interconnection strategy: Interconnect resources from at least two Chinese cloud providers, optimizing the network architecture across data centers (and even across cities)
Kratsios Accuses Without Unpublished Technical Evidence
According to the original text, Michael Kratsios, director of the Office of Technology Policy at the White House, made the following accusations on X: the U.S. government “has information showing” that Moonshot AI distilled Claude Fable 5 to develop K3; Moonshot AI built a precise internal platform that can quickly switch between multiple access methods to evade detection; Moonshot AI has already obtained Thailand-based servers equipped with GB300.
However, so far Kratsios has not published any technical evidence to support the above claims. The original text also clearly states that “this is still a one-sided government accusation, not a confirmed investigation conclusion.” In addition, some technical experts noted that Claude Fable 5 was only made public on July 1, and with K3’s release just three weeks later, it is technically impossible to use distillation as the main development path.
FAQ
What are the main differences between Moonshot AI’s K3 and K4?
According to Moonshot AI’s public statement, K3 was open-sourced on July 27, 2026, with 2.8 trillion parameters. K4 is still in the planning stage; its scale will be “clearly larger” than K3, and the company is actively seeking more Blackwell Hash Power.
How does Moonshot AI address the problem of insufficient Blackwell Hashrate?
According to the original text, Moonshot AI’s approach is to combine Hashrate from at least two Chinese cloud providers and have engineers optimize the network architecture across data centers (even across cities) so that distributed training can run smoothly.
Has Kratsios provided technical corroboration for his accusations against Moonshot AI?
According to the original text, Kratsios’s accusations (distilling Claude Fable 5, obtaining GB300-equipped servers in Thailand) have not yet been accompanied by technical evidence to support them. The original text also clearly states that “this is still a one-sided government accusation, not a confirmed investigation conclusion.”
