DeepSeek V4 Pro VRAM requirements
DeepSeek V4 Pro needs about 993 GB of VRAM at Q4_K_M with 8K context, so it needs a 1,128 GB setup such as 8× H200; a 24 GB card is too small. DeepSeek V4 Pro has 1.60T parameters, of which about 49B are used per token; all 384 experts still have to be in memory. With the same 8K context and one request it needs 1,639 GB at FP8 and 3,277 GB at FP16/BF16. The published weights take 805 GB (4.3 bits/weight). It does not fit on a 24 GB card at Q4_K_M; the smallest setup here that holds it with an 8K context is 8× H200 (1,128 GB).
993 GB at Q4_K_M, 8K context, one request
- FP8
- 1,639 GB
- FP16 / BF16
- 3,277 GB
- Published weights
- 805 GB
- Smallest setup, Q4_K_M
- 8× H200 (1,128 GB)
Worked example: DeepSeek V4 Pro with 32K tokens
Inputs: DeepSeek V4 Pro · Q4_K_M weights · 32,768 tokens of context · one request · FP16 KV cache
- Weights (1.60T at Q4_K_M)
- 901 GB
- KV cache (122 KB per token)
- 3.81 GB
- Buffers and runtime (0.5 GB + 10%)
- 91.0 GB
- Total
- 996 GB
The smallest setup here that holds it is 8× H200 (1,128 GB): split across the 8 cards, with the 0.5 GB runtime base on each of them, it takes 999 GB and leaves about 129 GB to spare. Change the inputs in the calculator
What makes DeepSeek V4 Pro's memory use different
As a mixture-of-experts model it reads about 27.6 GB of its 901 GB Q4_K_M weights per generated token (3%), so it writes like a much smaller model while needing memory for all of them.
How much VRAM does DeepSeek V4 Pro need?
DeepSeek V4 Pro needs 993 GB at Q4_K_M, 1,742 GB at Q8_0 and 3,277 GB at FP16 with 8,192 tokens of context and one request. Each total below is the weights plus the FP16 KV cache and the runtime overhead (0.5 GB plus 10%), and each row opens the calculator with that setting. What the GGUF names mean.
| Precision | Weights | Total | Smallest setup |
|---|---|---|---|
| As published (4.3 bits/weight) | 805 GB | 887 GB | 8× H200 (1,128 GB) |
| FP16 / BF16 | 2,978 GB | 3,277 GB | None listed |
| FP8 / INT8 | 1,489 GB | 1,639 GB | None listed |
| INT4 (AWQ / GPTQ) | 791 GB | 872 GB | 8× H200 (1,128 GB) |
| GGUF Q8_0 | 1,582 GB | 1,742 GB | None listed |
| GGUF Q6_K | 1,221 GB | 1,345 GB | None listed |
| GGUF Q5_K_M | 1,055 GB | 1,162 GB | None listed |
| GGUF Q4_K_M | 901 GB | 993 GB | 8× H200 (1,128 GB) |
| GGUF IQ4_XS | 810 GB | 892 GB | 8× H200 (1,128 GB) |
| GGUF Q3_K_M | 728 GB | 802 GB | 8× H200 (1,128 GB) |
| GGUF IQ3_XXS | 614 GB | 677 GB | 8× H200 (1,128 GB) |
| GGUF Q2_K | 624 GB | 687 GB | 8× H200 (1,128 GB) |
Fine-tuning DeepSeek V4 Pro? DeepSeek V4 Pro VRAM for LoRA, QLoRA and full training.
DeepSeek V4 Pro at 8K, 32K, 128K and 1M (full) tokens of context
The tables use an independent-layer upper-bound proxy for its 61-layer shared, compressed cache. Each extra token of context adds 122 KB of FP16 cache per request. The last column is the smallest setup here, one card or a group, that holds it at Q4_K_M. How the KV cache works.
| Context | KV cache, FP16 | Total at Q4_K_M | Total at Q8_0 | Smallest setup, Q4_K_M |
|---|---|---|---|---|
| 8K tokens | 976 MB | 993 GB | 1,742 GB | 8× H200 (1,128 GB) |
| 32K tokens | 3.81 GB | 996 GB | 1,745 GB | 8× H200 (1,128 GB) |
| 128K tokens | 15.3 GB | 1,008 GB | 1,758 GB | 8× H200 (1,128 GB) |
| 1M tokens | 122 GB | 1,126 GB | 1,875 GB | None listed |
How fast DeepSeek V4 Pro writes on a multi-GPU group
Tokens per second for one request with 8,192 tokens of context, estimated from memory bandwidth and the parameters read per token. No single card or machine holds it at Q4_K_M or FP8 (the precisions below), so these rows are the GPU groups here that do, in tensor parallel with their combined bandwidth. A dash means it does not fit at that precision. Try other settings in the speed calculator.
| Hardware | Bandwidth | Q4_K_M | FP8 |
|---|---|---|---|
| 8× H200 | 38,400 GB/s | 162–345 | — |
Longest context on a multi-GPU group
No single card or machine holds it at Q4_K_M, Q8_0 or FP8 (the precisions below), so these rows use GPU groups: how many tokens of context DeepSeek V4 Pro fits on each, in one tensor-parallel group with one request, an FP16 KV cache and 0.5 GB left free. "Full" means the model's whole context window fits.
| Group | Memory | Q4_K_M | Q8_0 | FP8 |
|---|---|---|---|---|
| 2× RTX 3060 12GB | 24 GB | No | No | No |
| 2× RTX 3090 | 48 GB | No | No | No |
| 2× RTX 4090 | 48 GB | No | No | No |
| 2× RTX 5090 | 64 GB | No | No | No |
| 4× RTX 3090 | 96 GB | No | No | No |
| 8× H100 SXM | 640 GB | No | No | No |
| 8× H200 | 1,128 GB | 1011K | No | No |
Model details
- Parameters
- 1.60T (1,598,839,674,782)
- Experts
- 384 routed experts, all loaded
- Active per token
- 49B
- Layers
- The tables use an independent-layer upper-bound proxy for its 61-layer shared, compressed cache
- Attention cache
- 1 KV heads × 512
- Context length
- 1,048,576 tokens
- Published weights
- 805 GB (4.3 bits/weight)
- On Hugging Face
- deepseek-ai/
DeepSeek-V4-Pro
Why this estimate looks this way
At Q4_K_M and an 8K-token context, DeepSeek V4 Pro uses 901 GB for weights, 976 MB for its FP16 KV cache and 90.7 GB for estimated runtime overhead, totaling 993 GB. The overhead is a 0.5 GB base plus 10% of weights and cache.
This is a mixture-of-experts model: all 384 experts contribute to the 1.60T parameters held in memory, even though only about 49B parameters run per token. Using only active parameters would understate VRAM.
This model shares and compresses its KV cache across layers, which the calculator does not model, so the KV cache figure is an upper bound.
The architecture and published checkpoint size come from the model's config.json and weight files. These are estimates rather than measured peak memory; inference engines can reserve extra buffers or preallocate the full configured cache.
DeepSeek V4 Pro next to similar-sized models
The three models here closest to it in parameter count, at Q4_K_M with 8,192 tokens of context.
| Model | Parameters | Total | KV per token | Smallest setup |
|---|---|---|---|---|
| DeepSeek V4 Pro | 1.60T, 49B active | 993 GB | 122 KB | 8× H200 (1,128 GB) |
| Qwen3.8 2.4T-A95B | 2.45T, 95B active | 1,517 GB | 92 KB | None listed |
| MiMo V2.6 Pro | 1.02T, 42B active | 636 GB | 50 KB | 8× H100 SXM (640 GB) |
| Kimi K3 | 2.78T, 104B active | 1,724 GB | 27 KB | None listed |
DeepSeek V4 Pro VRAM questions
How much VRAM do I need to run DeepSeek V4 Pro locally?
About 993 GB at Q4_K_M with an 8K context and one request, by the bits-per-weight estimate, and 3,277 GB at BF16. The smallest setup listed here that holds Q4_K_M is the 8× H200 (1,128 GB).
How much more VRAM does DeepSeek V4 Pro need for 32K or 128K tokens of context?
Its FP16 KV cache is 976 MB at 8K, 3.81 GB at 32K, 15.3 GB at 128K for one request, so 128K adds 14.3 GB over 8K, plus 10% overhead. A q8_0 cache
Does DeepSeek V4 Pro need less VRAM because only 49B parameters are active?
No. All 384 experts stay loaded, so the Q4_K_M weights take 901 GB. The active parameters decide speed: each generated token reads about 27.6 GB of them.
Badge for your model card
Paste it into a Hugging Face or GitHub README; it shows the Q4_K_M total above and links to this page. Q8_0 and FP16 badges.
Numbers read from the model files on Hugging Face on . Compare it with other models in the reproducible model and GPU report.