- Real-time web search across 100+ websites
- Analyze up to 50 files (PDFs, Docs, PPTs, Images) with ease
- AI slides & websites maker
- State of the art coding capabilities
- Enhanced image understanding beyond basic text extraction
This is the 8th launch from Kimi AI - Now with K2.5. View more
Kimi K2.5
Launched this week
Kimi K2.5 is Kimi's most intelligent model to date, achieving open-source SoTA performance in Agent, code, visual understanding, and a range of general intelligent tasks. It is also Kimi's most versatile model to date, featuring a native multimodal architecture that supports both visual and text input, thinking and non-thinking modes, and dialogue and Agent tasks.
Free
Launch Team
Flowtica Scribe
Hi everyone!
Multi-agent architectures are evolving, and Kimi K2.5 executes the "Swarm" concept at a level of scale and native integration we haven't seen before.
Instead of just making a single model think longer (scaling up), they are scaling out.
K2.5 introduces the "Agent Swarm" paradigm. For complex tasks, it autonomously spawns up to 100 sub-agents to execute workflows in parallel, reducing execution time by up to 4.5x.
The Native Multimodality also looks practical, especially the ability to generate code directly from screen recordings (Video-to-Code), rather than just static images.
Kimi also released Kimi Code, which integrates these agentic capabilities directly into the terminal and IDEs like @VS Code, @Cursor, @JetBrains and @Zed.
Impressive to see this level of capabilityâespecially the "Swarm" orchestrationâbeing open-sourced!
100 sub-agents sounds wild until you hit the coordination ceiling. Research keeps showing teams above 3-4 agents see communication overhead spike faster than reasoning gains. Would love to know if K2.5's orchestrator prunes inactive agents dynamically or runs more fire-and-forget.
Kimi K2.5 beats Opus 4.5 on every coding benchmark!? Wow.
FWIW the model is free for a week on @Kilo Code.
Product Hunt
@curiouskitty It's so true that the scaffold makes all the difference for agents! Itâs interesting how things like retry policies and sandboxes often do the heavy lifting compared to the base model. Getting those environment pieces right first seems like the smartest move for any team.
Can we use it to connect local databases or internal APIs to the Swarm for enterprise-level automation?
JDoodle.ai
Agent Swarm sounds really interesting. Does each agent on the swarm can make their own tool/mcp callings or they are consolidated by the main agent?