Kimi K3 · released 2026-07 · 2,780B parameters (104B active per token)
Hybrid KimiLinear: 69 Kimi-Delta-Attention layers (96 heads x 128, no KV cache) + 24 gated MLA layers (MLA cache 576 elements/token/layer). Latent MoE: routed experts run in a 3584-dim latent. Card: 2.8T total / 104B active, MoonViT-V2 401M included. KDA state assumed BF16.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.32 GB | 0.27 GB |
| 8,192 tokens | 0.43 GB | 0.32 GB |
| 32,768 tokens | 1.06 GB | 0.64 GB |
| 131,072 tokens | 3.59 GB | 1.90 GB |
| 1,048,576 tokens | 27.2 GB | 13.7 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 5,178.0 GB |
| FP8 (W8A8) | 8 | 2,593.4 GB |
| INT8 (W8A8 / W8A16) | 8 | 2,593.4 GB |
| NVFP4 | 4.5 | 1,462.6 GB |
| INT4 AWQ / GPTQ | 4.25 | 1,381.8 GB |
| GGUF Q8_0 | 8.5 | 2,750.8 GB |
| GGUF Q6_K | 6.56 | 2,123.0 GB |
| GGUF Q5_K_M | 5.69 | 1,841.4 GB |
| GGUF Q4_K_M | 4.89 | 1,582.5 GB |
| GGUF Q3_K_M | 3.91 | 1,265.4 GB |
| GGUF Q2_K | 3.35 | 1,084.1 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) | > 16 | > 16 |
| H200 SXM 141GBFP8 (W8A8) | > 16 | > 16 |
| H100 SXM 80GBFP8 (W8A8) | > 16 | > 16 |
| Instinct MI300X 192GBFP8 (W8A8) | > 16 | > 16 |
| A100 SXM 80GBINT4 AWQ / GPTQ | > 16 | > 16 |
| L40S 48GBFP8 (W8A8) | > 16 | > 16 |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | > 16 | > 16 |
| 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 |