Kimi K2 Thinking
moonshotai · released 2025-11-06 · other license
Kimi K2 Thinking is Moonshot's 1T-parameter (32B active) open-weight reasoning agent (Nov 2025), leading open models on agentic and reasoning benchmarks. Artificial Analysis rates it #1 among comparable open models (index 41).
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 1058.12B total · MoE, — active |
| Architecture | kimi_k2 |
| Context window | 262K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 84.5% |
|---|---|
| SWE-bench Verified | 71.3% |
| AIME | 94.5% |
| MMLU-Pro | 84.6% |
| BFCL v3 (tool use) | — |
| Composite score | 7.4 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 615.2 GB | 613.7 GB | — | 262K |
Strengths & weaknesses
Strengths: SOTA open-weight agentic reasoning with stable long-horizon tool use; Strong HLE-with-tools and agentic search; Native INT4, 256K context
Weaknesses: Verbose / high token use; Text-only, no image input