Chinese AI model Kimi k2.5 launches with parallel agent swarm and large context window
Moonshot AI has introduced Kimi K2.5, an open weight language model focused on scalable agent coordination. Its main addition is Agent Swarm, which can run up to 100 sub agents in parallel on complex tasks, support up to 1,500 tool calls per job, and reportedly cut execution time by as much as 4.5 times compared to a single agent.
K2.5 builds on Kimi K2 with a Mixture of Experts architecture featuring one trillion total parameters, 384 experts, and eight selected per token. It supports a 256,000 token context window, was trained on about 15 trillion tokens, and uses the MoonViT vision encoder for image and video understanding.
Moonshot AI also introduced Parallel Agent Reinforcement Learning, where a trainable orchestrator assigns subtasks to specialized agents and uses staged rewards to encourage parallelism without sacrificing quality. The company positions K2.5 as strong for coding, especially frontend work, and says it outperforms GPT-5.2 and Gemini 3 Pro on several agentic and multimodal benchmarks. K2.5 is available via Kimi.com, the mobile app, an API, and as downloadable weights on Hugging Face.
- AI Chatbot
- Freemium
- Proprietary
Related news
External links
- Kimi K2.5: Visual Agentic Intelligence Kimi Blog • Official source
- Moonshot AI releases open-source Kimi K2.5 model with 1T parametersSiliconANGLE
- China's Moonshot releases a new open source model Kimi K2.5 and a coding agentTechCrunch
- Moonshot's Kimi K2.5 introduces agent swarm, highlights open source model momentumConstellation Research
- Moonshot AI releases Kimi K2.5, claims most powerful open-weight model with 100-agent coordinationThe Decoder
Comments
Progress around "swarms" (which are "pools", but AI salesmen need to rename every possible thing to get funds) is interesting but very very costly, so comparing with one "agent" (again, let's use this misleading word) in terms of execution time (instead of money) is not really relevant. Also speedups often only appear on very lengthy tasks, often mistaking as "complex" tasks, that often don't provide very much more completion than shorter ones.
Also, Kimi compares cost effectiveness against ChatGPT, which is unfair giving how serious OpenAI is when it comes to burn money. And I don't even mention the "Kimi represents a meaningful step toward AGI" on its webpage, that at least, makes me smile.
The good thing is that it's open weight, so if anyone have been able to borrow some thousands untapped GB300 from Microsoft and is the owner of a medium-sized nuclear power plant, he can use this model for free.