DeepSeek V4 · released 2026-09 · 763B parameters (16.0B active per token)
Causal Encoder-Decoder (20 encoder + 20 decoder layers; decoder global KV projected from encoder). Card: 552B backbone + 196B Engram conditional-memory tables (+ vision encoder, 3 MTP/DSpark layers) = 763B safetensors; 8B active in prefill, 16B in decode. CSA2 sparse attention, FP4 KV cache.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.00 GB | 0.00 GB |
| 8,192 tokens | 0.01 GB | 0.01 GB |
| 32,768 tokens | 0.03 GB | 0.03 GB |
| 131,072 tokens | 0.11 GB | 0.11 GB |
| 1,048,576 tokens | 0.87 GB | 0.87 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 1,421.6 GB |
| FP8 (W8A8) | 8 | 713.3 GB |
| INT8 (W8A8 / W8A16) | 8 | 713.3 GB |
| NVFP4 | 4.5 | 403.4 GB |
| INT4 AWQ / GPTQ | 4.25 | 381.2 GB |
| GGUF Q8_0 | 8.5 | 755.2 GB |
| GGUF Q6_K | 6.56 | 582.8 GB |
| GGUF Q5_K_M | 5.69 | 505.5 GB |
| GGUF Q4_K_M | 4.89 | 434.5 GB |
| GGUF Q3_K_M | 3.91 | 347.4 GB |
| GGUF Q2_K | 3.35 | 297.6 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× · 74.7 tok/s | 8× · 51.6 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 8× · 56.4 tok/s | 8× · 36.1 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 16× · 26.4 tok/s | 16× · 21.6 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 8× · 59.9 tok/s | 8× · 38.8 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 8× · 48.7 tok/s | 8× · 30.1 tok/s |
| L40S 48GBFP8 (W8A8) | > 16 | > 16 |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 16× · 21.5 tok/s | 16× · 16.0 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 |