kimi-k2.5 moonshotai multimodality model-training mixture-of-experts agentic-ai vision video-understanding model-optimization parallel-processing office-productivity
MoonshotAI's Kimi K2.5 is a 32B active-1T parameter open-weights model featuring native multimodality with image and video understanding, built through continual pretraining on 15 trillion mixed visual and text tokens. It introduces a new MoonViT vision encoder and supports advanced capabilities like Agent Swarm, which coordinates up to 100 sub-agents for parallel workflows, and an Office Productivity K2.5 Agent for large-scale office tasks. This release marks a significant leap in open models from China, claiming state-of-the-art results on benchmarks like HLE and BrowseComp, and offering aggressive API pricing and throughput.