Kimi K2.5 Lands on Ollama with Native Multimodal Agentic Capabilities — 15T Token Training, Swarm Execution, Vision-Grounded Tool Use

Ollama·medium signal

Moonshot AI's Kimi K2.5, an open-source multimodal agentic model trained on ~15 trillion mixed visual and text tokens, is now available on Ollama (171K+ stars). K2.5 supports native vision-language understanding with agentic tool use grounded in visual inputs, code generation from visual specifications, and a swarm-like multi-agent execution scheme. The model follows K2.6 (April 20) but represents the first open-weight model to ship native agent swarm capabilities for local execution — significant for developers wanting multi-agent orchestration without cloud APIs.