“The training for Kimi K3 involved million-token agentic reinforcement learning with persistent rollout and sandbox states.”