A
argbe.tech - news
1min read

Kimi K2.5 adds image input and multi-agent tool orchestration

Moonshot’s Kimi K2.5 expands the K2 line from text-only to multimodal and promotes a built-in agent swarm mode for parallel tool use.

Moonshot’s Kimi K2.5 launched today with image input support and an emphasis on parallel, tool-using agents.

  • The earlier Kimi K2 model arrived in July 2025 as an open-weight 1T-parameter LLM; Kimi K2 Thinking followed in November 2025 with added reasoning.
  • K2.5 is described as natively multimodal and trained with roughly 15T mixed visual+text tokens.
  • The model advertises automatic agent swarm execution: up to 100 sub-agents running parallel work across as many as 1,500 tool calls, with a claimed speedup of up to 4.5× versus single-agent execution.
  • A public demo shared via OpenRouter’s chat UI had K2.5 generate an SVG illustration from a text prompt.
  • The Hugging Face release is listed at about 595GB, and its modified MIT-style license adds UI attribution requirements above 100M monthly active users or $20M monthly revenue.