Kimi K2
Moonshot AI’s flagship coding model. 1 trillion parameter MoE with 32B active parameters. Achieves 65.8% on SWE-bench Verified at 1/10th the cost of Claude Opus 4.
Specifications
| Spec | Value |
|---|---|
| Total Parameters | 1 trillion |
| Active Parameters | 32 billion |
| Experts | 384 total, 8 selected per token |
| Training Data | 15.5T tokens |
| Context Window | 128K tokens |
Performance
Benchmark Scores
| Benchmark | Score | Notes |
|---|---|---|
| SWE-bench Verified | 65.8% | Behind Claude 4, ahead of most |
| LiveCodeBench | 53.7% | Coding tasks |
| MATH-500 | 97.4% | Math problems |
| Output Speed | 47.1 tokens/sec | Fast inference |
| First Token Latency | 0.53s | Quick response |
Real-World Examples
Django Bug Fix:
- Identified validation logic in 3 files
- Implemented fix with error handling
- All tests passed first attempt
- Time: 12 seconds, Cost: $0.02
Agentic Capabilities
- Native MCP Support: Model Context Protocol
- Multi-step Reasoning: Trained on simulated tool interactions
- Code Execution: Write, debug, iterate autonomously
- Task Decomposition: Breaks complex problems into steps
Pricing Comparison
| Model | Input/M | Output/M | Monthly (100M tokens) |
|---|---|---|---|
| Kimi K2 | $0.15 | $2.50 | $15 |
| Claude Opus 4 | $15 | $75 | $1,500 |
| GPT-5 | $2.50 | $10 | $250 |
Savings: 100x cheaper than Claude Opus 4
Access
- API: platform.moonshot.ai
- OpenAI-compatible
- HuggingFace: MoonshotAI/Kimi-K2-Instruct (MIT license)
- Block-fp8 format weights
Notes
- Purpose-built for software engineering
- Optimized for agentic workflows
- Part of “Chinese AI Trinity” (Kimi, Qwen, GLM)
[2026-07-20] 后继型号
2026-07-17 Moonshot 发布 Kimi K3（2.8T 参数 / 1M 上下文 / 智能指数 57 / 约 Opus 4.8 一半成本），定位从”高性价比编码模型”升级为”接近前沿的通用智能体模型”，并催化全球算力股重定价。K2 作为前代旗舰，见 Kimi K3。
Related: Moonshot-AI | Kimi-K3 | China-AI-Model-Landscape-2025-2026 | AI-Models-Landscape-2025-2026