Kimi K2.5: Unified Multimodal Intelligence and Agent Swarms скачать в хорошем качестве

Kimi K2.5: Unified Multimodal Intelligence and Agent Swarms

Describes the architecture and performance of the Kimi K2.5 model, as well as the innovative Agent Swarm system. This model integrates images and high-resolution videos through the MoonViT-3D encoder, and efficiently solves complex challenges through *parallel agent reinforcement learning (PARL)**. In particular, the orchestrator dynamically creates sub-agents and distributes tasks in parallel, reducing processing time and maximizing context management capabilities. In addition, the Toggle training technique was introduced to increase the efficiency of the operation by reducing the output length while maintaining the reasoning performance. As a result, Kimi K2.5 demonstrated **global top (SOTA)* performance in various indicators such as mathematical reasoning, coding, and long-term video analysis. Finally, show the model's powerful tool utilization and practical agent performance through real-world gameplay analysis or complex visual puzzle solutions. https://arxiv.org/pdf/2602.02276