Kimi K2 Instruct is Moonshot AI’s latest open-source model designed for agentic intelligence—it doesn’t just answer, it acts. Built with a 1T parameter MoE architecture, it achieves state-of-the-art performance in reasoning, code generation, and real-world task execution.
Kimi K2 is reflex-grade, designed to understand tools, environments, and user goals—without needing complex workflows.
Key Features
- Agentic by Design: Executes tasks across tools and environments with no manual orchestration.
- State-of-the-Art MoE Model: 1T total parameters, 32B active per forward pass.
- Reflex-Speed Decisions: No long thinking loops—optimized for fast, smart action.
- Open Access: Comes in both base and instruct versions, supporting custom fine-tuning.
- Multi-Tool Support: Interacts across tools like calendars, files, browsers, terminals, emails, and APIs.
- Terminal-Native: Runs and debugs CLI tasks, edits files, and executes system commands.
- High-Performance Benchmarks: Outperforms open and closed models on agentic, reasoning, and coding tasks.
- No Workflow Scripting Required: Just describe your task—Kimi K2 handles the rest.
Agentic Use Cases
- Data Insights: Analyze and visualize salary data with 16 IPython calls.
- Web Agents: Explore Stanford NLP Genealogy using web search, scrolls, edits, and deployments.
- Trip Planning: Books flights, stays, and meals for a 2025 Coldplay tour—all autonomously.
- Code Transformation: Converts a Flask project to Rust with benchmarking and validation.
- Game Automation: Builds and debugs Minecraft JavaScript mods with test case management.
- Model Analysis: Uses Weights & Biases to interpret logs and write reports.
Training & Architecture
- Pretraining: Built with high token-efficiency and a focus on tractable exploration.
- Post-training: Refined via reinforcement learning on large-scale agentic simulations.
- Agentic Simulation Data: Generated across domains and toolsets with LLM-evaluated rubrics.
- Inspired by ACEBench: Simulates realistic, multi-turn tool-use scenarios for robust RL learning.
Model Variants
- Kimi-K2-Instruct: Pretrained and post-trained for general-purpose instruction and reflex-grade action.
- Kimi-K2-Base: Raw foundation model for researchers and developers to build on.