Moonshot shipped Kimi K3.1 Pro on October 23, with the API up immediately and full weights promised by November 4. At the same moment, Beijing's MOFCOM and MOST published a draft AI export catalogue that covers models at 2 trillion parameters or more, and Kimi's reported 2.8 trillion total parameters put it squarely in the band. The release date and the comment deadline for the export rules land in the same two weeks. Whether the weights actually ship freely depends on which side of that line the rule falls on.
The model itself is not the surprising piece. Moonshot priced it at roughly 2 dollars input and 10 dollars output per million tokens, and the coding benchmark scores put it just behind the top closed frontier. What is the story is the cyber posture. Moonshot said it stripped cyber-related data out of mid-training and filtered the cyber reinforcement-learning environments before the open-weight release, deliberately capping the model's offensive capability at about 81 percent pass on its internal hard benchmark versus 90 percent for a checkpoint that did not make the cuts. The company is signaling it does not want to extend the frontier of hacking tools, in the same week the state is drafting rules about exactly that.
The second-order effect is who gets to download the weights. Officials said API and SaaS access does not count as an export, but transferring weights or source code abroad does. That distinction is going to matter for the November 4 weight drop: a single download by an overseas developer is a very different legal act than running the model on someone else's cloud. The gap between the promised release date and the November 15 comment window is where the real uncertainty lives, and it will shape how fast the next open Chinese frontier model reaches the rest of the world.
Image source: i.ibb.co