Kimi K2 Instruct
moonshotai · released 2025-07-11 · other license
Kimi K2 Instruct is Moonshot's July 2025 1T-parameter (32B active) non-thinking open-weight model for chat and agentic tool use, with 65.8% SWE-bench Verified single-attempt and strong math/tool-use scores (vendor).
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 1026.47B total · MoE, — active |
| Architecture | kimi_k2 |
| Context window | 131K tokens |
| Knowledge cutoff | 2024-12-31 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 75.1% |
|---|---|
| SWE-bench Verified | 65.8% |
| AIME | 49.5% |
| MMLU-Pro | 81.1% |
| BFCL v3 (tool use) | — |
| Composite score | 6.89 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 596.9 GB | 595.4 GB | — | 131K |
Strengths & weaknesses
Strengths: Strong non-thinking agentic and tool-use model; Excellent math (MATH-500 97.4) and instruction following (IFEval 89.8); 1T MoE, open weights
Weaknesses: No long chain-of-thought; AIME 2025 only 49.5; Weak terminal/agentic coding vs thinking models