Quick answer
How do I run DeepSeek V4 on a cloud GPU?
To run DeepSeek V4 on a cloud GPU, rent enough VRAM for its weights plus about 20%: V4 Flash (284B parameters, 159.6 GB of weights) needs at least two H200s, and the SGLang cookbook documents four; V4 Pro (1.6T, 864.7 GB) needs eight H200s. Serve it with vLLM or SGLang. As of 29 October 2026, GPUwithCrypto rents the H200 at $3.07/hour per GPU, paid in crypto.
DeepSeek V4 at a glance (Hugging Face, read 29 October 2026)
| Models | DeepSeek-V4-Flash and DeepSeek-V4-Pro, open weights under the MIT License |
|---|---|
| Parameters | Flash: 284B total, 13B activated per token · Pro: 1.6T total, 49B activated |
| Weight format | Mixed: MoE expert weights in FP4, most other weights in FP8 |
| Weights on disk | Flash 159.6 GB (46 files) · Pro 864.7 GB (64 files) |
| Context length | 1 million tokens; DeepSeek recommends at least 384K for the Think Max mode |
| Serving engines on the model card | vLLM and SGLang |
| Sampling advice on the card | temperature 1.0, top_p 1.0 |
How much VRAM does DeepSeek V4 need?
Flash needs about 191.5 GB and Pro about 1,037.6 GB: the published weights × 1.2. DeepSeek V4 is a mixture-of-experts model. Only 13B (Flash) or 49B (Pro) parameters work on each token, which cuts compute, not memory: every expert must sit in VRAM. Its compressed attention keeps the KV cache small, so the usual per-token cache formula for Llama-style models does not apply. The weights × 1.2 rule for other models is on how much VRAM do I need?
Which GPUs hold DeepSeek V4?
Formula minimum = the fewest GPUs whose combined VRAM holds weights × 1.2. Documented layouts = what the vLLM recipes and the SGLang cookbook describe. Our price per hour for the whole configuration, 29 October 2026.
| Model | Weights | × 1.2 | Formula minimum | Documented layouts | GPUwithCrypto per hour |
|---|---|---|---|---|---|
| V4 Flash | 159.6 GB | 191.5 GB | 2× H200 (282 GB) or 2× B200 (360 GB) | SGLang: 4 GPUs on H200 or B200, 8 on H100 · vLLM: 4 GPUs per replica on H200 or B200 | $6.14/hour; $12.28/hour for the documented layout |
| V4 Pro | 864.7 GB | 1,037.6 GB | 8× H200 (1,128 GB) or 8× B200 (1,440 GB) | SGLang: 8 GPUs on H200 or B200; H100 needs two 8-GPU nodes · vLLM: 8 GPUs on H200, context capped at 800K tokens | $24.56/hour |
Prices checked on
8× H100 (640 GB) is too small for V4 Pro, and our hosts have at most 8 GPUs. The H100 vs H200 page compares their memory and price. H100 and H200 have no FP4 Tensor Cores: SGLang runs the FP4 experts on them with W4A16 kernels, while the B200 runs FP4 natively.
How do I serve DeepSeek V4 Flash with vLLM?
The layout below is the one the vLLM recipe recommends: four GPUs, data parallel plus expert parallel. Every flag comes from that recipe.
Rent four H200s
Start a 4× H200 instance (564 GB). Check that its disk holds the 159.6 GB of weights plus room for logs. Every instance has local NVMe disk, sized for its GPU and shown before you start. The disk is wiped when you delete the instance, so copy your checkpoints out first.
Install vLLM
Install the latest tagged vLLM release; the recipe asks for vLLM 0.20.0 or newer for this model. You can also start the instance from the
vllm/vllm-openaiDocker image.Download the weights
hf download deepseek-ai/DeepSeek-V4-Flashsaves the 46 weight files to the Hugging Face cache. vLLM also downloads them on first start.Start the server
Run the command below. It binds to 127.0.0.1 so the API is reachable only through your SSH tunnel.
Send a request
Forward port 8000 over SSH and call the OpenAI-compatible endpoint
/v1/chat/completions.Stop the instance when you are done
Instances are billed per minute at the hourly rate while they run, from the GPU-hours you bought for that GPU. Stopping an instance stops the charge at once.
pip install -U vllm
hf download deepseek-ai/DeepSeek-V4-Flash
vllm serve deepseek-ai/DeepSeek-V4-Flash \
--trust-remote-code \
--data-parallel-size 4 \
--enable-expert-parallel \
--kv-cache-dtype fp8 \
--block-size 256 \
--tokenizer-mode deepseek_v4 \
--reasoning-parser deepseek_v4 \
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--host 127.0.0.1 --port 8000
# forward the API port (host and SSH port are shown in the dashboard)
ssh -N -L 8000:127.0.0.1:8000 -p SSH_PORT root@HOST
# in another terminal
curl http://127.0.0.1:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "deepseek-ai/DeepSeek-V4-Flash", "messages": [{"role": "user", "content": "Explain the KV cache in two sentences."}]}'
How do I serve DeepSeek V4 with SGLang instead?
Start the instance from the lmsysorg/sglang:latest Docker image, which the SGLang cookbook lists for H100, H200 and B200, and use tensor parallelism: --tp 4 for Flash on four H200s, --tp 8 for Flash on eight H100s or for Pro on eight H200s. The API answers on port 30000 with the same OpenAI-compatible routes.
python3 -m sglang.launch_server \
--model-path deepseek-ai/DeepSeek-V4-Flash \
--trust-remote-code \
--tp 4 \
--host 127.0.0.1 \
--port 30000
How do I serve DeepSeek V4 Pro?
On eight H200s, with data parallelism across all eight GPUs. The vLLM recipe for H200 caps the context at 800K tokens to leave room for the cache, because the dense weights are copied to every GPU. For the Think Max mode, keep at least 384K tokens of context, as DeepSeek recommends.
vllm serve deepseek-ai/DeepSeek-V4-Pro \
--trust-remote-code \
--data-parallel-size 8 \
--enable-expert-parallel \
--max-model-len 800000 \
--kv-cache-dtype fp8 \
--block-size 256 \
--tokenizer-mode deepseek_v4 \
--reasoning-parser deepseek_v4 \
--host 127.0.0.1 --port 8000
How much does it cost to run DeepSeek V4 per hour?
At our prices, V4 Flash on the documented four-H200 layout costs $12.28/hour, and V4 Pro on eight H200s costs $24.56/hour. A full day of one H200 is $73.68. Provider by provider: H200 rental prices.
Pro's weights are 5.4 times the size of Flash's (864.7 against 159.6 GB). According to GPUwithCrypto's price check of 29 October 2026, the lowest self-serve on-demand H200 price was $3.62 per GPU-hour (Massed Compute): $14.48 per hour for four GPUs and $28.96 for eight. Every price: pricing.
Which DeepSeek V4 checkpoint should I download?
Start with deepseek-ai/DeepSeek-V4-Flash or deepseek-ai/DeepSeek-V4-Pro, the repositories created on 22 April 2026, whose cards give the figures above. DeepSeek later published updated checkpoints, DeepSeek-V4-Flash-0731 and DeepSeek-V4-Pro-0813, with a built-in draft head for speculative decoding; the vLLM recipe asks for vLLM 0.25.0 or newer for them. The -Base repositories are for further pre-training, not for chat or tool calls.
DeepSeek V4 questions
How much VRAM do I need for DeepSeek V4 Flash?
About 191.5 GB: 159.6 GB of weights plus 20%. Two H200s (282 GB) hold it by that rule; the SGLang cookbook and the vLLM recipe both document four H200s, which leaves more room for the cache and for parallel users.
How much VRAM do I need for DeepSeek V4 Pro?
About 1,037.6 GB: 864.7 GB of weights plus 20%. Eight H200s (1,128 GB) or eight B200s (1,440 GB) hold it on one host. Eight H100s (640 GB) do not.
Can I run DeepSeek V4 on H100s?
V4 Flash, yes: the SGLang cookbook documents eight H100s with --tp 8. V4 Pro needs two 8-GPU H100 nodes in the cookbook, more than one of our instances holds, so use eight H200s.
Can I run DeepSeek V4 on one GPU?
Not on the GPUs we rent. The largest, a 180 GB B200, holds Flash's 159.6 GB of weights but leaves only 20 GB for the cache, below the 20% margin. Two GPUs is the floor.
Can I pay for the H200s with crypto?
Yes. GPUwithCrypto takes BTC, USDT, USDC, ETH, SOL, LTC from a $25 top-up, with no ID and no card. GPU-hours are sold in blocks of 10, 24, 168 or 720 hours for one GPU model, at the same hourly rate; instances of that GPU draw from them per minute, and unused hours never expire.
Sources (8)
- DeepSeek-V4-Flash model card and files — DeepSeek / Hugging Face (read 2026-10-29)
- DeepSeek-V4-Pro model card and files — DeepSeek / Hugging Face (read 2026-10-29)
- vLLM recipe: DeepSeek V4 Flash — vLLM project (read 2026-10-29)
- vLLM recipe: DeepSeek V4 Pro — vLLM project (read 2026-10-29)
- DeepSeek-V4 cookbook — SGLang (read 2026-10-29)
- Hugging Face CLI guide (hf download) — Hugging Face (read 2026-10-29)
- NVIDIA H200 specifications — NVIDIA (read 2026-10-29)
- Massed Compute pricing vm.massedcompute.com/pricing — Massed Compute (read 2026-10-29)