Leaked API Parameters Reveal Moonshot AI's Upcoming Kimi K3.1 Model
Leaked internal API responses indicate Chinese AI startup Moonshot AI is preparing to launch Kimi K3.1, reportedly scheduled for next month. The model leak details support for three reasoning intensity levels (Low, High, Max), a 1-million-token context window, and native agent features like Swarm multi-agent collaboration. If verified, Kimi K3.1 brings system-level multi-agent orchestration and flexible reasoning compute directly into the API layer. This aligns with broader industry trends toward agentic workflows and offers developers built-in infrastructure for complex, autonomous AI operations. Discovered API codes highlight specific identifier models such as 'k3d1-agent', along with specialized processing modes for search and batch tasks. K3.1 builds on Kimi K3's architecture, which incorporates 2.8 trillion parameters, sparse MoE, Kimi Delta Attention, and Attention Residuals.
## BACKGROUND
Moonshot AI is a leading Chinese LLM developer recognized for its long-context Kimi model series. Its flagship Kimi K3 architecture leveraged technical innovations such as Kimi Delta Attention (KDA)—a linear attention mechanism using fine-grained gating to reduce KV cache footprint—and Attention Residuals (AttnRes), which replaces standard residual connections with learned attention across layers.