Moonshot: kimi-k2.6

Kimi K2.6 is a native multimodal, trillion-parameter mixture-of-experts (MoE) model engineered for state-of-the-art agentic reasoning and long-horizon task execution. Building on the foundation of K2.5, it expands the "Agent Swarm" architecture to support up to 300 parallel sub-agents coordinating over 4,000 tool-call steps. It features a dual-mode system: "Instant Mode" for rapid dialogue and "Thinking Mode" for deep reasoning, capable of autonomous runs lasting over 12 hours. In 2026, Kimi K2.6 is recognized as the premier open-weight alternative for complex software engineering, multi-day deep research, and large-scale document orchestration.