DeepSeek, China · Released April 22, 2026
DeepSeek V4 Pro has 1598.8B parameters (49B active per token). In BF16 its weights alone take 2930 GiB; quantized to 4-bit, about 885 GiB. Each 32K-token request adds 0.28 GiB of KV cache. Cheapest way to run it today: 5 × AMD MI300X (192 GB) at $11.95 per hour.
Size your own setup in the calculator →
DeepSeek, China · Released April 22, 2026ollama run deepseek-v4-pro 0 pulls (opens in a new tab)Weights plus the KV cache of one request (BF16 cache) and the model's fixed recurrent state, in GiB. Add 1 to 3 GiB for the inference engine itself.
| Weights format | Weights only | + 8K context | + 32K context | + 128K context | + 1M context |
|---|---|---|---|---|---|
| FP4 experts + FP8 (native) | 792 | 792 | 793 | 793 | 801 |
| BF16 / FP16 | 2930 | 2930 | 2930 | 2931 | 2939 |
| FP8 / INT8 | 1467 | 1467 | 1467 | 1468 | 1475 |
| 4-bit (GGUF Q4_K_M) | 885 | 885 | 885 | 886 | 893 |
Exact size of each quantization in DevQuasar/deepseek-ai.DeepSeek-V4-Pro-GGUF (opens in a new tab), and the smallest GPU that runs it with llama.cpp, Ollama or LM Studio and an 8K-token context.
| Quantization | File size | Smallest GPU |
|---|---|---|
| Q5_K_M | 1039 GiB | More than 96 GB: see the calculator |
| Q4_K_M | 886 GiB | More than 96 GB: see the calculator |
| Q3_K_M | 697 GiB | More than 96 GB: see the calculator |
| Q2_K | 530 GiB | More than 96 GB: see the calculator |
Read every day from Hugging Face. Split files are added together; vision projectors are left out.
Fewest GPUs that fit the weights plus one 8K-token request. Data-center GPUs use vLLM (90% of memory usable), the others llama.cpp. Several GPUs split the model with tensor or pipeline parallelism.
| GPU | FP4 experts + FP8 (native) | BF16 / FP16 | FP8 / INT8 | 4-bit (GGUF Q4_K_M) |
|---|---|---|---|---|
| GeForce RTX 4090 24 GB | 48 (several servers) | > 128 | 96 (several servers) | 48 (several servers) |
| GeForce RTX 5090 32 GB | 32 (several servers) | 128 (several servers) | 64 (several servers) | 32 (several servers) |
| RTX PRO 6000 Blackwell 96 GB | 16 (several servers) | 48 (several servers) | 24 (several servers) | 16 (several servers) |
| Mac M5 Max 128 GB 96 GB usable | 16 (several servers) | 48 (several servers) | 24 (several servers) | 16 (several servers) |
| NVIDIA H100 80 GB | 16 (several servers) | 48 (several servers) | 24 (several servers) | 16 (several servers) |
| NVIDIA H200 141 GB | 7 | 24 (several servers) | 16 (several servers) | 8 |
| NVIDIA B200 (HGX) 180 GB | 6 | 24 (several servers) | 16 (several servers) | 6 |
For each GPU: the fewest cards that fit DeepSeek V4 Pro with one 8K-token request, the most faithful weight format at that count, and the cheapest on-demand price (2026-10-10).
| Setup | Per hour | Weights | Provider | Per month | Rent |
|---|---|---|---|---|---|
| 5 × AMD MI300X 192 GB | $11.95 | FP4 experts + FP8 (official) | RunPod | $8,723.50 | Rent 5 × AMD MI300X on RunPod (opens in a new tab) → |
| 4 × AMD MI355X / MI350X 288 GB | $21.96 | FP4 experts + FP8 (official) | RunPod | $16,030.80 | Rent 4 × AMD MI355X / MI350X on RunPod (opens in a new tab) → |
| 7 × NVIDIA H200 141 GB | $32.25 | FP4 experts + FP8 (official) | Vast.ai | $23,541.77 | Rent 7 × NVIDIA H200 on Vast.ai (opens in a new tab) → |
| 4 × NVIDIA B300 / GB300 288 GB | $35.96 | FP4 experts + FP8 (official) | RunPod | $26,250.80 | Rent 4 × NVIDIA B300 / GB300 on RunPod (opens in a new tab) → |
| 6 × NVIDIA B200 (HGX) 180 GB | $37.51 | FP4 experts + FP8 (official) | Vast.ai | $27,379.38 | Rent 6 × NVIDIA B200 (HGX) on Vast.ai (opens in a new tab) → |
On-demand prices from RunPod, Vast.ai, Verda and Azure, checked every hour. Serving many users needs more KV cache, so more memory: size it in the calculator. Some links are affiliate links: StudioTV may earn a commission, at no extra cost to you.
Architecture: MoE 384 experts · 512-wide latent: 128-token window + sequence compression (30 layers 4×, 31 layers 128×). Its KV cache is 6.8× smaller than if every layer used full attention, because it keeps only a short window of tokens at full detail and compresses the rest of the sequence. Only 49B parameters are active per token, but all 1599B must sit in memory: VRAM depends on the total, speed on the active part.
| Context per request | BF16 cache | FP8 cache |
|---|---|---|
| 8K tokens | 0.08 GiB | 0.05 GiB |
| 32K tokens | 0.28 GiB | 0.17 GiB |
| 128K tokens | 1.10 GiB | 0.68 GiB |
| 1M tokens | 8.72 GiB | 5.39 GiB |
Plus 17.8 MiB of recurrent state per request, whatever the context.
In BF16 the weights alone take 2930 GiB (3146 GB). In FP8 that is 1467 GiB, and about 885 GiB with 4-bit quantization (Q4_K_M); the official FP4 experts + FP8 checkpoint is 792 GiB. Each request then adds KV cache: 0.08 GiB at 8K tokens and 0.28 GiB at 32K (BF16 cache).
No. Even in 4-bit it needs 48 (several servers) RTX 4090s. In 4-bit it needs 16 (several servers) H100 GPUs.
On 2026-10-10, the cheapest on-demand setup is 5 × AMD MI300X (192 GB) with FP4 experts + FP8 weights on RunPod, at $11.95 per hour (about $8,723.50 per month). Next: 4 × AMD MI355X / MI350X (288 GB) with FP4 experts + FP8 weights on RunPod, at $21.96 per hour. Sized for one 8K-token request; prices are checked every hour.
With one request and an 8K-token context: 48 (several servers) in BF16, 24 (several servers) in FP8 and 16 (several servers) in 4-bit (vLLM, 90% of memory usable). Serving many users at once needs more memory for their KV caches: the calculator sizes that for you.
For a model card or a README: the 4-bit size and the smallest GPU it fits on, linked to this page. Data as JSON: /api/models/deepseek-v4-pro.json.
[](https://studiotvai.com/vram-requirements/deepseek-v4-pro)Same family (large moe): MiniMax M2.7 · DeepSeek R1 · Kimi K2.6 · GLM-5.3 · GLM-5.3-Flash · DeepSeek V4 Flash · DeepSeek V4.1 Flash · MiMo-V2.6 Pro · MiMo-V2.6 Flash · Kimi K3 · Mistral Large 4 · DeepSeek V4 Flash Vision Exp · Qwen3.8 2.4T-A95B · Ornith 1.5 397B · Atria Dawn Preview · Intern S2 397B
All models: Llama 3.1 8B · Mistral Small 3.2 24B · Qwen3 32B · Llama 3.3 70B · Llama 3.1 405B · Qwen3.5 4B · Qwen3.5 9B · Qwen3.8 27B · Qwen3.6 35B-A3B · Qwen3.5 122B-A10B · Qwen3.8-Flash-Next · LFM2.5 8B-A1B · LFM2 24B-A2B · Gemma 4 12B · Gemma 4 31B · Gemma 4 26B-A4B · gpt-oss 20B · gpt-oss 120B · Llama 4 Scout 17B-16E · Llama 4 Maverick 17B-128E · Kolibri-1 · Nemotron 3 Nano 30B-A3B · Nemotron 3 Super 120B-A12B · MiniMax M2.7 · DeepSeek R1 · Kimi K2.6 · GLM-5.3 · GLM-5.3-Flash · DeepSeek V4 Flash · DeepSeek V4.1 Flash · MiMo-V2.6 Pro · MiMo-V2.6 Flash · Kimi K3 · Mistral Large 4 · Qwen3-Coder-Next 80B-A3B · Qwen3-Coder 30B-A3B · Qwen3-Coder 480B-A35B · DeepSeek V4 Flash Vision Exp · Qwen3.8 2.4T-A95B · Ornith 1.5 397B · Ornith 1.5 35B-A3B · Ornith 1.5 9B · Atria Dawn Preview · Intern S2 397B · Clef · Clef Flash · D1 3B · Humanizer · JEV 27B VL · Mellum2.1 12B-A2.5B · LightOnOCR-3 4B · Spark-X2.5 4B · Agnes-3.0-Qwen
Estimates, not guarantees: computed from the official config.json with the same engine as the LLM VRAM Calculator. Real usage depends on your engine version and settings.