Kimi-K2.5

Kimi K2.5 is built upon Kimi K2, a trillion-parameter mixture-of-experts transformer pre-trained on 15 trillion tokens, giving the model a massive foundation of linguistic and visual knowledge to draw from. Rather than treating vision and language as separate pipelines that compete for model capacity, K2.5 employs joint optimization so that the two modalities reinforce each other—visual understanding deepens language reasoning, and language comprehension sharpens visual interpretation. This unified architecture enables the model to handle design-to-code tasks, visual reasoning, and complex multimodal problems with a coherence that earlier separate-modality systems struggled to achieve.

The model introduces Agent Swarm, a parallel orchestration framework that decomposes complex problems into heterogeneous sub-tasks and dispatches them to up to 100 specialized agents running simultaneously. This self-directed coordination slashes execution latency by up to 4.5x compared to sequential single-agent approaches, while achieving 50.2% on Humanity's Last Exam at roughly 76% lower cost than comparable proprietary models. Post-trained through continual pretraining on roughly 15 trillion mixed visual and text tokens, Kimi K2.5 blends instant and thinking modes with both conversational and agentic paradigms, making it suitable for production-grade agentic workflows, autonomous coding, and real-time multimodal reasoning tasks.