More Manus than OpenAI: Quick Test of Kimi's New Multi-Core Agent Cluster
Discover Kimi K2.5, an all-in-one AI model with 256K context, image and video understanding, and dynamic Agent clusters. Outperforms OpenAI and Manus with 80% faster runtime and 400% efficiency boost
“Top Python Libraries” Publication 500 Subscriptions 20% Discount Offer Link.
Kimi has been holding back a massive trump card.
Instead of heavily relying on Claude like other general-purpose Agents, it first focused on building a truly Agentic model. The result? They created an All-in-One Kimi K2.5 model that can understand images and videos, has a 256K context window, and, in Agent cluster mode, can stably summon 100+ clones (sub-agents).
Unlike Manus’s Wide Research — where the clones all do the same type of task — these sub-agents can simultaneously handle completely different kinds of work.
For a long time, the general-purpose Agents I’ve used, like OpenAI’s Deep Research, felt like dealing with a very smart PhD student. You give it a research topic, it spends a huge amount of time diving deep, and finally delivers a detailed report.
Then Manus introduced Wide Research: it’s like dispatching 100 interns at once, all doing the exact same thing — e.g., gathering information — and then aggregating the results. This approach is extremely efficient for tasks that require massive information collection and initial sorting.
But with Kimi’s Agent cluster this time, I see a brand-new possibility. The dozens or even hundreds of Agents it deploys are no longer interns all doing the same job — they form a clearly divided team with specialized roles. And the team size dynamically adjusts based on task difficulty.
