Original video in Chinese.
Key Takeaways
- K2.5 brings vision, text, conversation, agent calls, and reasoning modes into one open model. The launch comparisons discussed here showed strong results against GPT-5.2-xhigh at a fraction of its cost.
- Agent Swarm divides work and uses dynamic coordination points to exchange information and align results, an approach I call three-dimensional orchestration.
- I see the shift from an individual model to organized intelligence as significant. At the time of this episode, Agent Swarm was in beta for members.
Human civilization depends on organization: division of labor, collaboration, and scale. I think the same principle applies to machine intelligence.
We have seen models with huge context windows and deep reasoning. These are still individual “geniuses” competing on individual capability.
Yesterday, Kimi launched K2.5 and Agent Swarm, taking its AI toward organized intelligence.
Why K2.5 combines so many capabilities
The word I would use for K2.5 is versatile. It combines vision and text, conversation and agent calls, thinking and non-thinking modes. It also performed strongly in the launch evaluations discussed here.
It is an open model. In several comparisons, it exceeded GPT-5.2-xhigh at a fraction of the cost, putting further pressure on OpenAI’s position.
Why put so many capabilities into one model? My interpretation is that a versatile model is needed to drive a system that creates and coordinates many different agents.
Looking back, I see three stages in Kimi’s development.
The first, around Kimi 1.0, emphasized memory and context length. The second, from 1.5 to 2.0, emphasized reasoning and tool use, especially in long tasks.
Both improved the individual model. The third emphasizes collaboration: AI organizing other AI. K2.5 is the model behind that stage.
Example one: a large literature sweep
I asked Kimi to search major academic and technical repositories and identify and catalog papers, reports, and official documents.
The task required subject expertise to find the right materials and substantial parallel capacity for the volume.
Kimi created four subagents covering different channels, including arXiv and GitHub. Each received detailed requirements such as date ranges, authors, and topic keywords.
They worked simultaneously. I could see progress below the main interface and open details. Completion appeared as a notification and a full progress bar.
Once all four finished, the main agent integrated their findings into a final report, including visual analysis. That involved both writing documents and code, which suited K2.5’s range of abilities.
The deliverables were saved in the virtual machine: dozens of substantial references and a detailed report. A human team would have been needed to complete that much work in a short period.
This illustrates the possibility of exchanging compute cost for expensive specialist time. Experts in many fields command high hourly rates. Agents can assist with collection and analysis on demand.
The Agent Swarm capability described in this episode supported as many as 1,500 tool-call rounds for demanding tasks.
Example two: analyzing an animated film
I asked Kimi for a shot-by-shot analysis of the first battle in Mobile Suit Gundam Hathaway.
It created four specialist agents: battle, visual, composition, and identity analysts. The roles showed that it had decomposed the task into meaningful perspectives.
They worked in parallel, and the main agent integrated their analyses. The report covered battle pacing, audiovisual language, and the director’s intent, which I particularly appreciated.
Text alone was not enough, so I asked for a table-based report with hand-drawn storyboard illustrations. The versatile model could call image capabilities when needed, producing a style I liked.
What I mean by an agent swarm
First, roles emerge from the task. A fixed flowchart written by programmers limits the system to predefined jobs. Many earlier multi-agent projects specified roles, processes, and rules in advance, with humans organizing the AI.
Kimi’s approach felt more like forming a specialist team for each assignment. One agent might read financial reports, another search news, and another write code. Roles followed demand, across animation and AI papers alike.
Second, I describe its coordination as three-dimensional. A system like Manus Wide Research resembles two-dimensional concurrency: 100 interns work in parallel on separate tasks, and their outputs are piled together. They divide up the work without collaborating.
Kimi’s swarm has dynamic relay points. Agents can pause at important stages, exchange information, calibrate their work, and continue. There is both specialization and alignment.
Complex, general tasks need roles and decomposition chosen for the situation. A single agent or a fixed multi-agent recipe may be insufficient.
For the user, I describe the result as a ByteDance in your pocket: an organized group rather than a smarter individual model. For machine intelligence, I see it as a step from isolated genius toward civilization.
Agent Swarm was in beta at the time of this episode. If you had a Kimi membership, I encouraged you to try it.
That is all for this episode. If you want to understand AI, become a more capable independent creator, and meet people with similar interests, join the newtype community. See you next time.