On July 15, Moonshot AI released a teaser video on Bilibili, X platform, and Xiaohongshu, officially building up momentum for the Kimi K3 model, with the new product expected to be launched soon. The number "3" briefly shown at around 4 seconds in the video has been interpreted by the outside world as an explicit hint at Kimi K3.
An anonymous model appears on Arena.ai, comparison tests have been leaked
According to public information, an anonymous model codenamed Kivine has quietly appeared on the Arena.ai evaluation platform, and the industry generally speculates that it is associated with Kimi K3. At the same time, netizens have already shared multiple comparison test videos of Kimi K3 against Claude Fable5, as well as horizontal comparisons with the GPT-5.6 Sol model, showing that Moonshot AI has directly targeted the top-tier models with the strongest coding capabilities.
K2 series iterates rapidly, K2.6 supports up to 300 sub-Agents running in parallel
Moonshot AI's Kimi model series has previously released multiple versions. The Kimi K2 released in July 2025 is an open-source MoE architecture model with a total parameter count of 1T and activated parameters of 32B, featuring strong code capabilities and general Agent task abilities. In April 2026, Moonshot AI released and open-sourced Kimi K2.6, which comprehensively upgraded code writing, long-range task execution, and Agent cluster capabilities, supporting the scheduling of up to 300 sub-Agents to process tasks in parallel. Shortly after the release of K2.6, Moonshot AI announced that the old version API of the K2 series would be discontinued on May 25, and users were advised to switch to the latest model.
The rapid iteration from K2 to K2.6, and now the pre-launch promotion of K3, shows that Moonshot AI is advancing its model upgrades at a fast pace. This time, K3 directly targets Claude Fable5 and GPT-5.6 Sol, indicating that domestic large models are no longer content with catching up, but are now directly challenging overseas top-tier models in the programming and Agent capability fields.
