How much VRAM does DeepSeek V4 Pro need?
DeepSeek V4 Pro needs 805 GB for its weights as released (FP4 + FP8), 806 GB with an 8K context and 807 GB with 32K.
- Parameters
- 1.6T
- Active per token
- 49B
- Longest context
- 1M
- Released
- Apr 2026
- License
- MIT
- Strongest at
- Chat, Coding, Agents, Docs
Memory at each precision
| Precision | Weights | With 8K | With 32K | With 128K |
|---|---|---|---|---|
| As released (FP4 + FP8) | 805 GB | 806 GB | 807 GB | 807 GB |
| 3-bit | 728 GB | 729 GB | 729 GB | 730 GB |
| 2-bit | 559 GB | 560 GB | 560 GB | 561 GB |
| 1.6-bit | 307 GB | 308 GB | 309 GB | 309 GB |
One conversation, 16-bit context cache, runtime overhead included.
What runs it
| Machine | Best fit | Context | Speed | First reply |
|---|---|---|---|---|
| RTX 3060 12 GB | Too big | |||
| RTX 4060 Ti 16 GB | Too big | |||
| RTX 3090 | Too big | |||
| RTX 4090 | Too big | |||
| RTX 5090 | Too big | |||
| 2× RTX 3090 | Too big | |||
| RTX PRO 6000 Blackwell | Too big | |||
| M4 Max 128 GB | Too big | |||
| M5 Max 128 GB | Too big | |||
| Ryzen AI Max+ 395 128 GB | Too big | |||
| DGX Spark 128 GB | Too big | |||
| M3 Ultra 256 GB | Too big | |||
| M5 Ultra 512 GB | 1.6-bit | 1M | 47 tok/s | 9.5 s |
| H100 SXM | Too big | |||
| DGX Station GB300 | 2-bit, spills | 1M | 47 tok/s | 4.4 s |
| 8× H200 NVL | FP4 + FP8 | 1M | 64 tok/s | 0.4 s |
Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.
How to run it
- llama.cpp
llama-server -hf unsloth/DeepSeek-V4-Pro-0813-GGUF- vLLM
vllm serve deepseek-ai/DeepSeek-V4-Pro- MLX (Mac)
mlx_lm.chat --model mlx-community/DeepSeek-V4-Pro-4bit
Questions
How much VRAM does DeepSeek V4 Pro need?
DeepSeek V4 Pro needs 805 GB for its weights as released (FP4 + FP8), 806 GB with an 8K context and 807 GB with 32K.
Can DeepSeek V4 Pro run on an RTX 4090?
Not on one card. The smallest machine here that runs it is the M5 Ultra 512 GB.