Kimi K2

Kimi K2 is built on a mixture-of-experts architecture, meaning it routes problems through specialized subnetworks rather than activating the entire model at once. This design lets it punch far above its weight—32 billion active parameters backed by a trillion total parameters—while keeping inference costs low. The model is explicitly engineered for agentic work: it doesn't just answer queries, it takes action, invoking tools dynamically and autonomously executing complex, multi-step tasks. Open weights mean anyone can download and adapt it, and its benchmark results across software engineering, multilingual coding, and mathematics show it competing with models that cost far more to operate.

The model ships in two forms: a foundation version for researchers who want full control for fine-tuning, and an instruction-tuned variant ready for drop-in use in chat and agentic applications. The instruction variant is reflex-grade, optimized for speed without requiring long thinking processes. Its tool-use capabilities let developers give it a set of functions and describe a goal; Kimi K2 automatically figures out how to chain calls together to get there. Compared to models like Claude Opus 4 that cost $15 per million input tokens, Kimi K2's pricing makes it attractive for production agentic pipelines at scale.