Kimi K2 · released 2026-04 · 1,027B parameters (32.9B active per token)
Card: same architecture as K2.5 (1T / 32B active, MoonViT ~0.4B). Kimi-K2.7-Code (Jun 2026) has the identical checkpoint layout. DeepSeek-V3 architecture with 384 experts and 64 heads; MLA cache 576 elements/token/layer.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.27 GB | 0.13 GB |
| 8,192 tokens | 0.54 GB | 0.27 GB |
| 32,768 tokens | 2.14 GB | 1.07 GB |
| 131,072 tokens | 8.58 GB | 4.29 GB |
| 262,144 tokens | 17.2 GB | 8.58 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 1,912.7 GB |
| FP8 (W8A8) | 8 | 960.7 GB |
| INT8 (W8A8 / W8A16) | 8 | 960.7 GB |
| NVFP4 | 4.5 | 544.2 GB |
| INT4 AWQ / GPTQ | 4.25 | 514.5 GB |
| GGUF Q8_0 | 8.5 | 1,016.1 GB |
| GGUF Q6_K | 6.56 | 784.2 GB |
| GGUF Q5_K_M | 5.69 | 680.2 GB |
| GGUF Q4_K_M | 4.89 | 584.6 GB |
| GGUF Q3_K_M | 3.91 | 467.4 GB |
| GGUF Q2_K | 3.35 | 400.5 GB |
One server, 8K context, FP8 on data center GPUs and GGUF Q4_K_M on consumer and Apple hardware.
| GPU | 1 user | 32 users |
|---|---|---|
| B200 SXM (HGX B200)FP8 (W8A8) | 8× · 89.4 tok/s | 8× · 37.2 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 16× · 21.5 tok/s | 16× · 15.7 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 16× · 21.3 tok/s | > 16 |
| Instinct MI300X 192GBFP8 (W8A8) | 8× · 85.1 tok/s | 8× · 28.2 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 8× · 79.8 tok/s | 16× · 13.5 tok/s |
| L40S 48GBFP8 (W8A8) | > 16 | > 16 |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 16× · 20.6 tok/s | 16× · 10.5 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | > 16 | > 16 |
| GeForce RTX 4090 24GBGGUF Q4_K_M | > 16 | > 16 |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | > 16 | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |