The next flagship from Moonshot AI seems to have been "spoiled" by the code itself. Multiple sources claim that Kimi K3.1 is quietly taking shape internally, with estimates suggesting it could officially launch as early as next month (October 2026), and it will offer three adjustable levels of reasoning intensity: Low, High, and Max.
The first clue came from internal JSON and API responses within Moonshot AI. Leaker @MaxForAI posted on X platform on September 23, revealing model identifiers like "k3d1-agent" from captured internal snippets. These included Agent mode, application scenarios, and several configuration switches. By following these configuration files, the capability map of K3.1 has become quite clear: it may support an ultra-long context option called "Extra Long," with a maximum of 1 million Tokens; native Agent mode; further introducing Swarm multi-Agent collaboration; and task modes such as search and batch processing are also listed. In other words, the model is no longer just a chat interface but is preparing to split itself into a team capable of division of labor, collaboration, and external tool integration.
Bringing the focus back to the current flagship, Kimi K3 was the strongest publicly released model from Moonshot AI in July 2026. According to official data, it has 2.8 trillion parameters, native visual capabilities, and a 1 million Token context window. Its architecture uses Kimi Delta Attention, Attention Residuals, and sparse MoE. If K3.1 truly leaks as per the configuration, upgrading the K3 base by splitting the reasoning intensity into three levels and making Agent and Swarm modes ready-to-use would mean evolving from "a single model working alone" to "a model leading a group of models." Users can choose Low for cost-saving or Max for higher quality based on task importance, or even delegate complex tasks directly to multiple agents for parallel processing.
Currently, these identifiers are still in the internal response and configuration file stages, and Moonshot AI has not officially announced them yet. The exact October release date is still just speculation. However, the code wouldn't randomly include "k3d1-agent" and Swarm switches. When a company writes Agent, ultra-long context, and multi-Agent collaboration into configurations at this density, it's clear that the target of the next round of large model competition has shifted from "who chats better" to "who can work independently."
