Moonshot's Kimi K3.1 leaks with 1M token context and multiple reasoning levels

Tech media Wccftech reports that Chinese AI company Moonshot is preparing to launch its Kimi K3.1 model, expected to arrive in October 2026. The model will offer three reasoning levels: Low, High, and Max.

A source on X discovered Kimi K3.1 identifiers in Moonshot's internal JSON/API responses, including model names like "k3d1-agent," along with references to Agent mode and various configuration switches.

According to the config files, K3.1 may support multiple reasoning levels, an extra-long context of up to 1 million tokens, Agent mode, Swarm multi-agent collaboration, and task modes for search and batch processing.

Kimi K3 launched in July 2026 with 2.8 trillion parameters, native vision capabilities, and a 1 million token context window, using architectures like Kimi Delta Attention, Attention Residuals, and sparse MoE.