Moonshot AI Launches Kimi K2.5 With Vision Capabilities and 100-Agent Parallel Task Execution, Rivaling GPT-5 and Gemini
Summary
Moonshot AI launches Kimi K2.5, a powerful open-weight multimodal AI capable of vision processing and running up to 100 parallel sub-agents simultaneously, with benchmark scores rivaling GPT-5 and Gemini, and is now publicly available via API, chat, and Hugging Face.
Key Points
- Moonshot AI has released Kimi K2.5, an open-weight multimodal LLM that adds vision capabilities to its predecessor Kimi K2, achieving benchmark scores comparable to frontier models like GPT-5 and Gemini.
- Kimi K2.5 introduces an Agent Swarm mode, powered by a new reinforcement learning technique called PARL, which allows the model to autonomously decompose complex tasks and execute them in parallel across up to 100 sub-agents.
- Benchmarks show Kimi K2.5 outperforms GPT-5.2 Pro on BrowseComp and Claude Opus 4.5 on WideSearch, with significant reductions in wall-clock time, and the model is now publicly available via chat interface, API, and Hugging Face.