Kimi K2
Last updated: April 22, 2026
Specs
| Vendor | Moonshot AI |
|---|
| Released | November 7, 2025 |
|---|
| Context window | 2M tokens |
|---|
| Input | $0.5 / M tokens |
|---|
| Output | $2 / M tokens |
|---|
Benchmarks
| mmlu | 89.1 |
|---|
| gpqa | 64.7 |
|---|
| humanEval | 90.1 |
|---|
| arenaElo | 1384 |
|---|
| sweBench | 65.8 |
|---|
| sweBenchPro | 56 |
|---|
| liveCodeBench | 70 |
|---|
| aiderPolyglot | 60 |
|---|
| terminalBench | 38 |
|---|
| mmluPro | 81 |
|---|
| hle | 12 |
|---|
| aime | 70 |
|---|
| math500 | 92 |
|---|
| mmmu | 68 |
|---|
Strengths
- Frontier-grade open model from Moonshot (China)
- 2M context window with reliable retrieval
- Strong on long-form writing and agentic workflows
Weaknesses
- Western hosting options still limited
- Tool-calling ecosystem less mature than OpenAI
- Support docs skewed to Chinese
Best for
- Open-weights teams needing 1M+ context
- APAC-facing products with Chinese and English
- Research groups wanting self-hosted long-context reasoning
Compare with other AI models