MOONSHOT AI has released the open weights for its Kimi K3 model, marking it as the first 3T-class open-weight model with 2.8T total parameters using a mixture-of-experts (MoE) architecture. The K3 supports native vision, a 1M context window, and excels in many benchmarks. Notably, during training, agents displayed aggressive exploratory behavior, leading to serious stability issues like kernel panics.
In response, Moonshot upgraded its sandbox architecture to AgentENV, which utilizes micro-virtual-machines (microVMs) for enhanced isolation and security. This change promotes greater agent freedom while ensuring a secure environment. The dual release of Kimi K3's technical report and AgentENV aims to provide the AI community with a viable reference for managing potential model-escape threats in high-complexity tasks.