Kimi K2 Dark Side of the Moon Open Source 2025.7
The world's first open source commercially available large model with trillions of parameters, a pioneer in Agentic Intelligence
🌙 What is Kimi K2?
Kimi K2 is the flagship large model open sourced by Moonshot AI on July 11, 2025. As the world's first open source and commercially available large-scale model with trillions of parameters, Kimi K2 organically integrates Agentic Intelligence (autonomous agent intelligence) with powerful tool calling capabilities for the first time, and is comparable to DeepSeek V3 in multiple benchmark tests.
trillion parameters
1 trillion total parameters, 384 expert models
Agent optimization
Agentic Intelligence, powerful tool call
Completely open source
Apache 2.0 license, available for commercial use
🤖 Technical architecture
Kimi K2 Instruct Main version
Using MoE sparse architecture, only 32B parameters are activated for each inference (8 + 1 shared expert are dynamically selected among 384 experts), significantly reducing the computational load.
MoE Architecture
Hybrid expert architecture, dynamic routing, efficient reasoning
1.8-bit quantization
Unsloth quantized version only costs 245GB, original version 1.1TB
Multi-framework support
Support vLLM, llama.cpp, Docker deployment
📊 Performance comparison
| Model | Total parameters | Activation parameters | Context | Open Source |
|---|---|---|---|---|
| Kimi K2 | 1 trillion | 32B | 128K | Yes |
| DeepSeek V3 | 685 billion | 37B | 128K | Yes |
| Llama 4 Maverick | 400 billion | 17B | 128K | Yes |
| Claude 3.5 Sonnet | Unpublished | Unpublished | 200K | No |
💡 Core Competencies
Code generation
Programming capabilities are close to Claude 3.7, supporting multiple languages
Tool call
Powerful Function Calling capabilities
Multiple rounds of dialogue
Excellent contextual understanding and conversational coherence
Reasoning ability
Complex logical reasoning and mathematical problem solving
Agent scenario
Autonomous planning, decision-making and execution of tasks
Chinese optimization
Native Chinese support, excellent performance in Chinese scenes
🚀 Quick start
Claude Code integration
# 在 Claude Code 中使用 Kimi K2 /model moonshotai/kimi-k2-instruct
Deploying with vLLM
pip install vllm from vllm import LLM, SamplingParams llm = LLM(model="moonshotai/Kimi-K2-Instruct") sampling_params = SamplingParams(temperature=0.7, max_tokens=1024) prompts = ["你好,请介绍一下你自己"] outputs = llm.generate(prompts, sampling_params) print(outputs[0].outputs[0].text)
Use llama.cpp (quantized version)
# 下载 1.8-bit 量化版(245GB) # 原版 1.1TB,量化后减少 80% ./llama-cli -m kimi-k2-instruct-Q2_K.gguf \ -p "你是一个有帮助的AI助手" \ --n-predict 512
🎯 Application scenarios
AI Agent Development
Build an intelligent agent that can make autonomous decisions and call tools
Code Assistant
Replace Claude Code backend and reduce costs
Enterprise Application
Apache 2.0 license, available for commercial deployment
Local deployment
The quantitative version supports running on consumer-grade hardware
🔗 Related Resources
- Hugging Face: moonshotai/Kimi-K2-Instruct
- Official website: kimi.ai
- Quantitative version: Unsloth 1.8-bit quantization