DeepSeek R1 · released 2025-01 · 684B parameters (37.6B active per token)
MLA: KV cache per token per layer = kv_lora_rank + qk_rope_head_dim = 576 elements (latent, absorbed). params = safetensors total incl. ~14B MTP module (671B main model). Native FP8 (block 128x128) checkpoint. Reasoning model on the V3 base.
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 |
| 163,840 tokens | 10.7 GB | 5.36 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 1,275.0 GB |
| FP8 (W8A8) | 8 | 640.9 GB |
| INT8 (W8A8 / W8A16) | 8 | 640.9 GB |
| NVFP4 | 4.5 | 363.5 GB |
| INT4 AWQ / GPTQ | 4.25 | 343.7 GB |
| GGUF Q8_0 | 8.5 | 677.3 GB |
| GGUF Q6_K | 6.56 | 522.7 GB |
| GGUF Q5_K_M | 5.69 | 453.4 GB |
| GGUF Q4_K_M | 4.89 | 389.7 GB |
| GGUF Q3_K_M | 3.91 | 311.6 GB |
| GGUF Q2_K | 3.35 | 266.9 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× · 86.2 tok/s | 8× · 39.9 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 8× · 79.3 tok/s | 8× · 28.5 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 16× · 21.1 tok/s | 16× · 14.6 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 4× · 103 tok/s | 8× · 30.5 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 8× · 74.6 tok/s | 8× · 22.0 tok/s |
| L40S 48GBFP8 (W8A8) | > 16 | > 16 |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 16× · 20.2 tok/s | 16× · 11.0 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 16× · 20.1 tok/s | > 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 |