Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model
3 months ago
- #AI
- #Open-Source
- #Multimodal
- Introduction of Kimi K2.5, the most powerful open-source model to date.
- Kimi K2.5 builds on Kimi K2 with pretraining over 15T mixed visual and text tokens.
- Features state-of-the-art coding and vision capabilities with a self-directed agent swarm paradigm.
- Can self-direct an agent swarm with up to 100 sub-agents, reducing execution time by up to 4.5x.
- Available via Kimi.com, Kimi App, API, and Kimi Code with 4 modes: K2.5 Instant, Thinking, Agent, and Agent Swarm (Beta).
- Strong performance in agentic benchmarks (HLE, BrowseComp, SWE-Verified) at a fraction of the cost.
- Excels in coding, especially front-end development, turning conversations into complete interfaces.
- Capable of image/video-to-code generation and visual debugging.
- Evaluated using Kimi Code Bench, showing improvements over K2 across task types.
- Kimi Code integrates with IDEs like VSCode, Cursor, Zed, and supports images/videos as inputs.
- Agent Swarm uses Parallel-Agent Reinforcement Learning (PARL) for efficient distributed execution.
- Reduces end-to-end runtime by 80% and enables more complex workloads.
- Handles high-density office work, producing documents, spreadsheets, PDFs, and slide decks.
- 59.3% and 24.3% improvements over K2 Thinking in expert productivity benchmarks.
- Represents a step toward AGI for the open-source community.