Calculating practical deepseek v4 hardware requirements is essential for engineering teams, ML architects, and enterprise infrastructure leads deploying DeepSeek AI's frontier models. DeepSeek V4 spans two distinct architectural tiers: the flagship DeepSeek V4 Pro (a 1.6 trillion parameter Mixture-of-Experts model with 49B active parameters per token step) and the high-throughput DeepSeek V4 Flash (a 284 billion parameter MoE model routing 13B active parameters). Both models feature native support for 1,048,576 tokens (1M context) and Multi-Head Latent Attention (MLA).
GPU Picks aggregates hardware specifications, weight memory footprints, and cloud cluster pricing according to our transparent editorial methodology. This guide provides the calculated VRAM formulas, precision memory footprints (FP16, native FP8, and 4-bit AWQ), multi-GPU node blueprints across NVIDIA H200 and Blackwell B200 servers, and verified cloud deployment options.
Key takeaways
- Two Distinct Model Footprints: DeepSeek V4 Pro contains 1,600B (1.6T) total parameters activating 49B parameters per token step, while DeepSeek V4 Flash contains 284B total parameters routing 13B active parameters.[source][source]
- V4 Pro Memory Floor: In native FP8 precision, DeepSeek V4 Pro requires between 900 GB and 960 GB of aggregate VRAM for weights and basic KV state, necessitating a minimum cluster of 8x H200 141GB (1,128 GB VRAM) or 8x B200 192GB (1,536 GB VRAM).[source][source]
- V4 Flash Memory Floor: In native FP8 precision, DeepSeek V4 Flash requires approximately 180 GB to 190 GB of VRAM, running on a 2x H200 141GB or 4x A100/H100 80GB configuration.[source]
- Multi-Head Latent Attention (MLA): DeepSeek V4 compresses the Key-Value cache into low-rank latent vectors, reducing KV cache memory consumption by over 70% compared to traditional Multi-Head Attention at long context lengths.[source][source]
- Interconnect Bandwidth Demands: Multi-node MoE routing across V4 Pro requires high-speed interconnects (3.2 Tbps NDR InfiniBand or RoCEv2) to prevent all-to-all communication bottlenecks during expert dispatch.[source][source]
Direct answer: how to calculate deepseek v4 hardware requirements
To determine the GPU capacity required for DeepSeek V4 models, apply the Mixture-of-Experts parameter residency formula:
Because all experts must remain resident in GPU high-bandwidth memory for instant per-token routing, total parameter count dictates base memory capacity, while active parameter count determines the compute throughput (TFLOPS) needed per generation step.
| Provider | On-demand $/hr | Spot $/hr | Availability |
|---|---|---|---|
| Lambda Cheapest | $3.49 | n/a | Low |
| RunPod | $4.59 | $3.59 | Medium |
| CoreWeave | $6.31 | $2.62 | High |
DeepSeek V4 hardware requirements by model tier
The table below outlines memory requirements, precision formats, context boundaries, and recommended cloud GPU server nodes for both DeepSeek V4 Pro and DeepSeek V4 Flash.
| Model Tier | Precision Format | Total Parameter Weight | Context Cache (128k - 1M) | Recommended Total VRAM | Production Server Blueprint |
|---|---|---|---|---|---|
| DeepSeek V4 Pro (1.6T) | Native FP8 | ~900 GB | ~60 GB - 180 GB | ~1,050 GB - 1,150 GB | 8x H200 141GB or 8x B200 192GB |
| DeepSeek V4 Pro (1.6T) | 4-Bit AWQ / GGUF | ~500 GB | ~45 GB - 120 GB | ~580 GB - 650 GB | 4x H200 141GB or 8x H100 80GB |
| DeepSeek V4 Pro (1.6T) | FP16 (Full Precision) | ~3,200 GB | ~120 GB - 350 GB | ~3,500 GB | 2x (8x H200) or 16x B200 Cluster |
| DeepSeek V4 Flash (284B) | Native FP8 | ~180 GB | ~15 GB - 40 GB | ~210 GB - 230 GB | 2x H200 141GB or 4x A100 80GB |
| DeepSeek V4 Flash (284B) | 4-Bit AWQ / GGUF | ~95 GB | ~12 GB - 30 GB | ~115 GB - 130 GB | 2x A100 80GB or 1x H200 141GB |
| DeepSeek V4 Flash (284B) | FP16 (Full Precision) | ~580 GB | ~30 GB - 80 GB | ~640 GB | 8x H100 80GB or 6x H200 141GB |
For DeepSeek's newest 552B multimodal checkpoint, review our dedicated DeepSeek V4.1 Flash GPU requirements guide. Dedicated sizing for the earlier 284B checkpoint is covered in our DeepSeek V4 Flash 0731 guide.
DeepSeek V4 Pro architecture and cluster sizing (1.6T MoE)
DeepSeek V4 Pro is designed for frontier reasoning, complex multi-step coding, and enterprise agent workflows. Its architecture introduces several hardware-intensive characteristics:
1. 1.6T Total weights with 49B active routing
DeepSeek V4 Pro scales total model parameters to 1.6 trillion across 88 transformer layers with 128 attention heads.[source] In each forward pass, the router dynamically selects experts representing 49B parameters. While compute latency reflects a 49B dense model, the full 1.6T parameter tensor set must sit in memory.
2. Multi-Head Latent Attention (MLA) memory dynamics
Standard Multi-Head Attention (MHA) creates massive Key-Value cache pressure at 1M context. DeepSeek V4 utilizes Multi-Head Latent Attention (MLA), which compresses Key and Value matrices into a shared low-rank latent space.[source] This compression allows an 8x H200 node to maintain context windows up to 128k tokens across concurrent user batches without exhausting memory.
3. DSpark speculative decoding
DeepSeek V4 Pro incorporates DSpark speculative decoding modules.[source] DSpark uses small speculative draft heads to predict multiple tokens ahead, which are verified in parallel by the 49B active expert backbone. This increases memory bandwidth utilization on high-speed HBM3e GPUs like the NVIDIA H200 (4.8 TB/s) and Blackwell B200 (8.0 TB/s).[source][source]
DeepSeek V4 Pro Serving Topologies:
┌─────────────────────────────────────────────────────────────┐
│ 8x NVIDIA H200 141GB SXM5 (1,128 GB Total VRAM) │
│ ├─ Model Weights (FP8): ~900 GB │
│ ├─ MLA KV Cache Pool: ~160 GB │
│ └─ CUDA & PagedAttention Buffer: ~68 GB │
└─────────────────────────────────────────────────────────────┘
OR
┌─────────────────────────────────────────────────────────────┐
│ 8x NVIDIA Blackwell B200 192GB (1,536 GB Total VRAM) │
│ ├─ Model Weights (FP8): ~900 GB │
│ ├─ MLA KV Cache Pool (Extended 1M Context): ~500 GB │
│ └─ CUDA & Runtime Buffer: ~136 GB │
└─────────────────────────────────────────────────────────────┘
DeepSeek V4 Flash hardware requirements (284B MoE)
For production applications requiring lower latency and cost-effective hosting, DeepSeek V4 Flash offers a 284B parameter footprint routing 13B active parameters per token.[source]
Hardware configuration tiers for V4 Flash:
- Dual H200 141GB Node: Provides 282 GB of aggregate VRAM. In native FP8, V4 Flash consumes ~180 GB for weights, leaving ~100 GB for high-concurrency KV cache buffers.[source][source]
- Quad A100 / H100 80GB Node: Provides 320 GB of aggregate VRAM across 4 GPUs with NVLink. This configuration supports FP8 execution with comfortable head-room for medium batch sizes.[source]
- 4-Bit Workstation Setup: In 4-bit AWQ quantization (~95 GB weights), V4 Flash can run on 5x RTX 4090 24GB GPUs (120 GB total) or a single Mac Studio with 128GB Unified Memory for local development and offline evaluation.[source]
Cloud GPU deployment options and cluster costs
Deploying DeepSeek V4 models requires choosing between enterprise dedicated clusters and flexible neocloud compute. Live rates and instance availability can be compared directly in our GPU Lookup tool.
1. Nebius
Nebius provides dedicated 8x H200 and 8x H100 bare-metal and virtual nodes connected with 3.2 Tbps NVIDIA NDR InfiniBand fabric. Nebius is well suited for DeepSeek V4 Pro deployments requiring low-latency all-to-all expert communication across nodes.[source]
2. Lambda Labs & CoreWeave
Lambda Labs and CoreWeave offer high-density 8x H100 and 8x H200 SXM5 instances with dedicated Kubernetes orchestration. These platforms support automated model caching and tensor parallel deployment via vLLM or SGLang.[source]
3. RunPod & Vast.ai
For DeepSeek V4 Flash, RunPod and Vast.ai provide on-demand dual-GPU and quad-GPU instances (such as 2x H200 or 4x A100) at flexible hourly spot and on-demand pricing.[source]
Who should self-host DeepSeek V4 and who should skip it
Who should self-host DeepSeek V4
- Enterprises with strict data residency rules: Organizations handling proprietary codebases, financial records, or patient data that cannot leave private VPC networks.
- High-volume reasoning pipelines: Engineering teams generating millions of tokens daily for autonomous code agents, mathematical verification, or automated research pipelines.
- Infrastructure teams optimizing cost at scale: Deployments where dedicated 8x H200 clusters operate at continuous high utilization, beating per-token API rates.[source]
Who should skip self-hosting
- Early-stage teams with variable traffic: Teams generating moderate token volumes achieve substantially lower overall costs using managed API endpoints.
- Single-GPU developer workstations: DeepSeek V4 Pro cannot execute on single consumer GPUs. Teams without access to multi-GPU clusters should use DeepSeek V4 Flash in 4-bit quantization or access the API directly.[source]
- Teams without high-speed interconnect infrastructure: Hosting V4 Pro across nodes connected by standard Ethernet causes severe communication latency during MoE all-to-all dispatch.[source]
Serving engine configuration: vLLM and SGLang
DeepSeek V4 models achieve optimal throughput when configured with Tensor Parallelism (TP) and Expert Parallelism (EP) inside optimized inference engines such as vLLM or SGLang.[source]
Below is a production startup command for serving DeepSeek V4 Pro on an 8x H200 141GB node using SGLang with native FP8 quantization and MLA caching:
python3 -m sglang.launch_server \
--model-path deepseek-ai/DeepSeek-V4-Pro \
--tp 8 \
--quantization fp8 \
--kv-cache-dtype fp8_e5m2 \
--context-length 131072 \
--mem-fraction-static 0.90 \
--enable-mla \
--host 0.0.0.0 \
--port 30000For DeepSeek V4 Flash on a 2x H200 node:
python3 -m sglang.launch_server \
--model-path deepseek-ai/DeepSeek-V4-Flash-0731 \
--tp 2 \
--quantization fp8 \
--context-length 65536 \
--mem-fraction-static 0.88 \
--host 0.0.0.0 \
--port 30000Can I run DeepSeek V4 Pro on a single GPU?
No. DeepSeek V4 Pro contains 1,600 billion total parameters. In native FP8 precision, the model weights alone require approximately 900 GB of VRAM. The largest single datacenter GPUs currently available (NVIDIA Blackwell B200 192GB and NVIDIA H200 141GB) cannot hold the model weights alone. A minimum of 8x H200 141GB or 8x B200 192GB GPUs is required for production serving.
How much VRAM is required for DeepSeek V4 Flash?
DeepSeek V4 Flash requires approximately 180 GB to 190 GB of VRAM in native FP8 precision. In 4-bit AWQ or GGUF quantization, the memory footprint drops to approximately 95 GB to 115 GB. It can be hosted on a dual H200 141GB node, a quad A100 80GB node, or a 5x RTX 4090 24GB workstation cluster.
Why does DeepSeek V4 use Multi-Head Latent Attention (MLA)?
Multi-Head Latent Attention (MLA) compresses the Key-Value cache into low-rank latent vectors instead of storing full-dimensional key and value heads for every token. This reduces KV cache memory consumption by more than 70%, allowing servers to handle 128k to 1M context windows with significantly higher concurrency before encountering out-of-memory errors.[source]
What is the difference between DeepSeek V4 Pro and DeepSeek V4 Flash?
DeepSeek V4 Pro is a 1.6T parameter flagship model activating 49B parameters per token, engineered for frontier reasoning, complex mathematics, and deep coding tasks.[source] DeepSeek V4 Flash is a 284B parameter model activating 13B parameters per token, optimized for high token throughput, cost efficiency, and fast conversational responses.[source] For multimodal workflows, DeepSeek also provides DeepSeek-V4.1-Flash, scaling to a 552B parameter asymmetric architecture.
What network interconnect is required for DeepSeek V4 Pro clusters?
For multi-node DeepSeek V4 Pro clusters, NVIDIA NDR InfiniBand (3.2 Tbps) or 800G RoCEv2 networking is strongly recommended. Because the Mixture-of-Experts routing mechanism passes activation tensors between all GPUs during each forward step, standard 100G Ethernet introduces severe inter-node latency bottlenecks.[source]
Sources and references
- DeepSeek AI: DeepSeek-V4 Model Overview and Open Inference Checkpoints[source]
- DeepSeek AI: DeepSeek-V4-Flash-0731 Model Card[source]
- DeepSeek AI: DeepSeek-V4 Architecture and MLA Repository[source]
- NVIDIA Corporation: NVIDIA H200 Tensor Core GPU Architecture Datasheet[source]
- NVIDIA Corporation: NVIDIA Blackwell B200 Tensor Core GPU Datasheet[source]
- SGLang Project: SGLang DeepSeek-V4 High-Throughput Serving Documentation[source]