Kimi K2.5 Open Source: Moonshot AI Adds Native Multimodality and Agent Swarm
Moonshot AI open-sourced Kimi K2.5 in January 2026, offering native multimodal understanding, tool use, and parallel Agent Swarm.
Kimi advances from long-text and code models to vision-driven agent systems, attempting to let a main agent dynamically create multiple sub-agents to complete open-ended tasks in parallel.
K2.5 is based on continual pre-training on approximately 15T image-text tokens, adopts a MoE architecture with 1T total parameters and 32B activated parameters, supports 256K context and native INT4.
Domestic open-source models are now competing simultaneously on multimodality, tool chains, and multi-agent orchestration, with model capabilities and runtime design becoming further coupled.
Parallel agents should be prioritized for decomposable and verifiable research and engineering tasks, with evidence and failures tracked for each sub-task.
Observe the cost amplification of Swarm, sub-task independence, visual tool reliability, and community deployment reproducibility.