Kimi K2.5

Kimi K2.5 is Moonshot AI's most powerful open-source model, released January 27, 2026. It is a native multimodal agentic model — not a vision-adapted language model, but one trained from the ground up on approximately 15 trillion mixed visual and text tokens atop the Kimi-K2-Base.

**Native multimodality**: K2.5 processes vision and language as a unified representation, not through adapters or bridges. This gives it strong cross-modal reasoning: it can read UI designs, interpret video workflows, and generate grounded code or action plans based on what it sees — all without the quality degradation of bridged architectures.

**Agent Swarm — coordinated parallel intelligence**: Kimi K2.5 introduces a novel Agent Swarm execution model that scales beyond single-agent reasoning. The model decomposes complex tasks into parallel sub-tasks and dynamically instantiates domain-specific sub-agents to execute them simultaneously. Up to 100 specialized agents can coordinate on a single task, enabling a class of problems that sequential reasoning cannot solve efficiently.

**Dual-mode operation**: - **Instant mode** — Fast, direct responses for standard queries and tool use - **Thinking mode** — Extended internal reasoning for hard problems, comparable to dedicated reasoning models

**Agentic capabilities**: - Visual coding: generates code from UI screenshots, video workflows, and design specs - Autonomous tool orchestration and multi-turn task execution - Long context support for sustained agent pipelines

Kimi K2.5 is open-source under a permissive license, available on Hugging Face at moonshotai/Kimi-K2.5, and accessible via NVIDIA NIM and other inference providers.