DeepSeek V4 · released 2026-04 · 291B parameters (13.0B active per token)
Card: 284B total (safetensors 290.9B incl. MTP), 13B activated. 2 sliding-only + 21 CSA + 20 HCA layers. Hybrid attention: every layer keeps a 128-token sliding window of ONE shared K=V vector (512); CSA layers add one compressed 512-dim entry per 4 tokens plus a 128-dim Lightning-Indexer key per 4 tokens; HCA layers add one 512-dim entry per 128 tokens (see kvCompression; ~10% of V3.2's KV at 1M per card). Manifold-constrained hyper-connections (x4). Released weights: FP4 MoE experts + FP8 elsewhere. V4-Flash-0731 update (304B safetensors) published.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.03 GB | 0.02 GB |
| 8,192 tokens | 0.06 GB | 0.03 GB |
| 32,768 tokens | 0.22 GB | 0.11 GB |
| 131,072 tokens | 0.85 GB | 0.42 GB |
| 1,048,576 tokens | 6.72 GB | 3.36 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 541.9 GB |
| FP8 (W8A8) | 8 | 272.9 GB |
| INT8 (W8A8 / W8A16) | 8 | 272.9 GB |
| NVFP4 | 4.5 | 155.3 GB |
| INT4 AWQ / GPTQ | 4.25 | 146.8 GB |
| GGUF Q8_0 | 8.5 | 287.9 GB |
| GGUF Q6_K | 6.56 | 222.2 GB |
| GGUF Q5_K_M | 5.69 | 192.7 GB |
| GGUF Q4_K_M | 4.89 | 165.6 GB |
| GGUF Q3_K_M | 3.91 | 132.4 GB |
| GGUF Q2_K | 3.35 | 113.5 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) | 2× · 221 tok/s | 2× · 47.9 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 4× · 171 tok/s | 4× · 50.7 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 8× · 114 tok/s | 8× · 52.7 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 2× · 175 tok/s | 2× · 32.9 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 4× · 156 tok/s | 4× · 41.0 tok/s |
| L40S 48GBFP8 (W8A8) | 8× · 46.9 tok/s | 8× · 16.0 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 4× · 74.5 tok/s | 4× · 17.9 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 8× · 52.5 tok/s | 8× · 30.2 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 16× · 28.2 tok/s | 16× · 19.9 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | > 16 | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |