Tools / LLM GPU calculator / DeepSeek-V3 / R1
DeepSeek-V3 / R1 VRAM requirements
GPU memory for the weights and KV cache at FP16, FP8 and 4-bit, and how many GPUs that takes. Figures from the model's config.json.
Short answer
DeepSeek-V3 / R1 needs about 1,492 GB of GPU memory at FP16/BF16, 746 GB at FP8 and 411 GB at 4-bit to serve one user at an 8K-token context. Its 671B parameters take 1,342 GB at 16-bit, and that conversation adds 0.6 GB of KV cache. It is a mixture-of-experts model: 37B parameters are active per token, but all 671B must be in GPU memory. On NVIDIA H100 80 GB GPUs that is more than one 8-GPU server (at least 19 GPUs) at FP16 and more than one 8-GPU server (at least 10 GPUs) at FP8; for a desktop, at 4-bit it needs two 512 GB Mac Studios with a layer split.
Updated . KV cache at FP16, plus 10% headroom for activations and the runtime.
GPU memory by precision
| Weights | Weight memory | 1 user, 8K context | 32 users, 8K context | 1 user, 128K context |
|---|---|---|---|---|
| FP16/BF16 | 1,342 GB | 1,492 GB | 1,512 GB | 1,501 GB |
| FP8 | 671 GB | 746 GB | 766 GB | 756 GB |
| 4-bit (INT4) | 369.1 GB | 411 GB | 431 GB | 420 GB |
Required memory = (weights + KV cache) ÷ 0.9. An FP8 KV cache halves the cache share; set it in the calculator.
How many GPUs for one user at 8K context
| GPU | Memory | FP16/BF16 | FP8 | 4-bit (INT4) |
|---|---|---|---|---|
| Data center · tensor parallel | ||||
| NVIDIA L40S | 48 GB | Multi-server | Multi-server | Multi-server |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | 96 GB | Multi-server | 8 GPUs · TP8 | 8 GPUs · TP8 |
| NVIDIA H100 SXM | 80 GB | Multi-server | Multi-server | 8 GPUs · TP8 |
| NVIDIA H100 NVL | 94 GB | Multi-server | 8 GPUs · TP8 | 8 GPUs · TP8 |
| NVIDIA H200 | 141 GB | Multi-server | 8 GPUs · TP8 | 4 GPUs · TP4 |
| NVIDIA B200 | 180 GB | Multi-server | 8 GPUs · TP8 | 4 GPUs · TP4 |
| NVIDIA B300 | 288 GB | 8 GPUs · TP8 | 4 GPUs · TP4 | 2 GPUs · TP2 |
| AMD Instinct MI300X | 192 GB | 8 GPUs · TP8 | 4 GPUs · TP4 | 4 GPUs · TP4 |
| AMD Instinct MI355X | 288 GB | 8 GPUs · TP8 | 4 GPUs · TP4 | 2 GPUs · TP2 |
| Workstation & consumer · layer split | ||||
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | 16 cards | 8 cards | 5 cards |
| NVIDIA GeForce RTX 3090 | 24 GB | 63 cards | 32 cards | 18 cards |
| NVIDIA GeForce RTX 4090 | 24 GB | 63 cards | 32 cards | 18 cards |
| NVIDIA GeForce RTX 5090 | 32 GB | 47 cards | 24 cards | 13 cards |
| Apple Mac Studio M3 Ultra / M5 Ultra | 96 GB (72 usable) | 21 Macs | 11 Macs | 6 Macs |
| Apple Mac Studio M3 Ultra / M5 Ultra | 256 GB (192 usable) | 8 Macs | 4 Macs | 3 Macs |
| Apple Mac Studio M5 Ultra | 512 GB (384 usable) | 4 Macs | 2 Macs | 2 Macs |
Data center GPUs are counted at a tensor-parallel size (TP) that divides the model’s attention heads, within one 8-GPU server. Workstation and consumer cards are counted for a layer split, which works with any number of cards. Mac Studio counts about 75% of unified memory as usable by the GPU. Capacity only: speed depends on memory bandwidth and the serving stack.
Memory is step one. Throughput, latency, power and the network decide the real build.
Size my actual workloadDeepSeek-V3 / R1 architecture
- Parameters
- 671B total, 37B active per token (mixture of experts)
- Layers with a KV cache
- 61
- Attention heads / KV heads
- 128 / 128
- Head dimension
- 192
- Attention
- Multi-head latent attention (latent 512 + RoPE key 64)
- Maximum context
- 163,840 tokens
- KV cache per token (FP16)
- 61 layers × (512 latent + 64 RoPE key) × 2 bytes = 70,272 bytes (0.07 MB) per token
Source: deepseek-ai/DeepSeek-V3 config.json on Hugging Face. Parameter counts are the Hub’s count of the repo’s safetensors files, or the model card’s figure.
Questions about DeepSeek-V3 / R1
How much VRAM does DeepSeek-V3 / R1 need?
About 1,492 GB at FP16/BF16, 746 GB at FP8 and 411 GB at 4-bit, for one user at an 8K-token context. That is 1,342 GB, 671 GB or 369.1 GB of weights, plus 0.6 GB of KV cache, plus 10% headroom for activations and the runtime. Serving 32 users at 8K each takes 1,512 GB at FP16 or 766 GB at FP8, because every conversation keeps its own KV cache.
How many H100 GPUs does DeepSeek-V3 / R1 need?
At FP16/BF16 it takes more than one 8-GPU server (at least 19 GPUs), for one user at 8K or for 32 users. At FP8, more than one 8-GPU server (at least 10 GPUs), for one user at 8K or for 32 users. One user at FP8 takes 8 GPUs (tensor parallel 8) on H200 (141 GB) and 8 GPUs (tensor parallel 8) on B200 (180 GB).
Can DeepSeek-V3 / R1 run on an RTX 4090 or RTX 5090?
At 4-bit it needs about 411 GB for one user at 8K: 18 RTX 4090s or 13 RTX 5090s, more cards than one workstation holds. At FP8 (746 GB) it takes 32 RTX 4090s or 24 RTX 5090s, more cards than one workstation holds. Several cards work with a layer split, as in llama.cpp, which divides the model's layers between them.
Can DeepSeek-V3 / R1 run on a Mac Studio?
macOS lets the GPU use about 75% of unified memory by default: about 72 GB of a 96 GB, 192 GB of a 256 GB and 384 GB of a 512 GB Mac Studio. At 4-bit DeepSeek-V3 / R1 needs 411 GB for one user at 8K, so it needs two 512 GB Mac Studios with a layer split; at FP8 (746 GB) it needs two 512 GB Mac Studios with a layer split.
How large is DeepSeek-V3 / R1's KV cache?
At FP16, 61 layers × (512 latent + 64 RoPE key) × 2 bytes = 70,272 bytes (0.07 MB) per token. Multi-head latent attention caches one compressed latent per layer instead of keys and values for each of its 128 heads. An 8K-token conversation holds 0.6 GB at FP16, and a 128K-token one holds 9.2 GB.
Does DeepSeek-V3 / R1 need memory for all its experts?
Yes. DeepSeek-V3 / R1 is a mixture-of-experts model with 671B parameters, of which about 37B are active for each token. Any token can be routed to any expert, so all of them must be in GPU memory: memory follows total parameters, speed follows active ones.
What is DeepSeek-V3 / R1's maximum context length?
163,840 tokens, according to its Hugging Face configuration. One user at that full length holds 12 GB of KV cache at FP16, for 758 GB in total with FP8 weights, or 752 GB with an FP8 KV cache as well.
Other models
- Llama 3.1 8B
- Llama 3.3 70B
- Llama 3.1 405B
- Llama 4 Scout
- Llama 4 Maverick
- Qwen3 8B
- Qwen3 14B
- Qwen3 32B
- Qwen3-30B-A3B
- Qwen3-235B-A22B
- Qwen2.5 32B
- Qwen2.5 72B
- gpt-oss-20b
- gpt-oss-120b
- Gemma 3 27B
- Mistral Small 3.2 24B
- Mixtral 8x7B
- Phi-4 14B
- GLM-5.3
- LFM2.5-24B-A2B
- Kolibri-1
Not listed? The LLM GPU calculator sizes any Hugging Face model from its ID, link or config.json.
Need to deploy private AI?
We design it, validate it, source it, deploy it and help run it.
Give us your workload, users, budget and timeline. Within 48 hours we'll give you a validated cloud/on-prem architecture, BOM, expected costs and sourcing options.