How to run DeepSeek-V4 on your own GPU server

What does it take to run DeepSeek-V4 on your own server?

To run DeepSeek-V4 on your own server, you need GPU memory for the whole model, every expert included, plus room for the KV cache, and a recent version of vLLM or SGLang. By our sizing rule, DeepSeek-V4-Flash needs about 200 GB of GPU memory and DeepSeek-V4-Pro about 1,072 GB. Flash fits on a single B300; Pro needs a full 8-GPU B200 or B300 server.

• DeepSeek-V4-Flash: from $3,836 a month on 4× RTX PRO 6000, or $4,019 on one B300 (DeepSeek-V4-Flash VRAM and servers).
• DeepSeek-V4-Pro: from $27,512 a month on 8× B200 (DeepSeek-V4-Pro VRAM and servers).
• Software: vLLM 0.25.0 or later, or SGLang, with the flags of the official recipes.

The weights are open, so the model you deploy stays the one you run. DeepSeek’s own API has retired V4-Flash and answers requests for it with DeepSeek-V4.1-Flash (DeepSeek API docs). We publish no benchmarks: the figures below are memory sizes and prices, not speeds.

DeepSeek-V4-Flash vs Pro: the two models

DeepSeek released a preview of V4 on 24 April 2026 (DeepSeek). The official checkpoints, DeepSeek-V4-Flash-0731 and DeepSeek-V4-Pro-0813, supersede it and add a speculative decoding module, DSpark. They are the checkpoints our catalog installs.

Sources, checked on 19 October 2026: the model cards of DeepSeek-V4-Flash, DeepSeek-V4-Pro, Flash-0731 and Pro-0813.

• Both are mixture-of-experts models. Each token goes through 13B parameters in Flash and 49B in Pro, but every expert must sit in GPU memory, so size them by total parameters.
• Long context is built in. A hybrid of Compressed Sparse Attention and Heavily Compressed Attention keeps the KV cache small: at 1M tokens, DeepSeek states that Pro needs 10% of the KV cache of DeepSeek-V3.2.
• Three reasoning efforts: low, high and max. For max, DeepSeek recommends a context window of at least 384K tokens.
• Flash or Pro? Flash’s cheapest server costs $3,836 a month against $27,512 for Pro. DeepSeek reports that Flash-0731 outperforms the Pro preview on its listed benchmarks; the comparison that matters is Flash-0731 against Pro-0813 on your own prompts.

DeepSeek-V4 hardware requirements: how much VRAM?

We size every model the same way: official weights × 1.2 for the KV cache and activations, with 92% of GPU memory counted as usable. It is the rule of our open model pages and of the deploy page:

DeepSeek-V4-Flash-0731: 167 GB × 1.2 ≈ 200 GB
DeepSeek-V4-Pro-0813:   893 GB × 1.2 ≈ 1,072 GB

vLLM’s recipes set similar floors: 200 GB of GPU memory for Flash-0731 and 960 GB for Pro-0813 (Flash recipe, Pro recipe). Our model pages list one serving option for each, vLLM or SGLang with the official FP4/FP8 weights; our catalog has no Ollama build of either. The smallest server of each GPU that holds each model, by our rule:

— = larger than the biggest server of that GPU. n/a = ROCm support for DeepSeek-V4 is not confirmed in our catalog yet, so our model pages offer it on NVIDIA GPUs only. No GeForce server holds either model: four RTX 5090 give 118 GB of usable memory.

The cheapest fit for Flash is 4× RTX PRO 6000 at $3,836, with the model split over PCIe; one B300 costs $4,019 and holds it alone. For Pro it is 8× B200 at $27,512. The 20% margin covers moderate contexts and a few requests at a time; how much GPU memory an LLM needs explains the rule, and the sizing helper applies it to any model.

Multi-GPU: what fits in one server, and what does not

A GPUCenter server holds up to 8 GPUs, and servers are not joined by InfiniBand or any other fabric, so a model must fit inside one server. In the 8-GPU B300, B200, H200 and H100 servers, NVSwitch connects every GPU to every other at full NVLink speed, 1.8 TB/s per GPU on Blackwell and 900 GB/s on Hopper, which tensor and expert parallelism rely on (see multi-GPU). The RTX PRO 6000 and L40S communicate over PCIe.

• Flash fits from one GPU to eight. One B300 avoids any split. vLLM’s recipe runs one copy of Flash on four GPUs of an 8-GPU H200, B200 or B300 server, and verifies the RTX PRO 6000 as an 8-card server.
• Pro fits two configurations: 8× B200 (1,325 GB usable) and 8× B300 (2,120 GB).
• Pro does not fit on 8× H100 (640 GB in total), 8× RTX PRO 6000 (768 GB) or 8× L40S (384 GB): the 893 GB of weights alone exceed them, and two servers cannot be combined.
• Quantization does not shrink Pro much. Its checkpoint stores 1.6T parameters in 893 GB, about 4.5 bits per parameter, because the experts are already FP4; vLLM lists NVIDIA’s NVFP4 build with the same 960 GB minimum. Community builds at lower bit widths exist (Hugging Face): they are third-party, not in our catalog, and their quality is yours to test.

Deploy DeepSeek-V4 with vLLM or SGLang

Both engines support DeepSeek-V4 and publish recipes for it: vLLM’s recipes and SGLang’s cookbook. Four points before you start:

• Version. The 0731 and 0813 checkpoints need vLLM 0.25.0 or later, the release that added DSpark; the April preview runs from 0.20.0.
• No chat template. The official checkpoints ship no Jinja chat template. In vLLM, --tokenizer-mode deepseek_v4 applies DeepSeek’s encoding, so the OpenAI-compatible endpoint works as usual.
• Flags differ by GPU. The commands below are for 8-GPU B200 and B300 servers. On the H200 and the RTX PRO 6000, vLLM’s recipe uses other flags, and on the RTX PRO 6000 it currently needs vLLM’s nightly image for Flash-0731 with DSpark.
• Bind to 127.0.0.1 and set an API key, then reach the API through an SSH tunnel.

Pro on 8× B200 or 8× B300 with vLLM, in the recipe’s default layout: tensor parallelism with expert parallelism across the eight GPUs.

KEY=$(openssl rand -hex 32); echo "$KEY"   # the API key for your clients
vllm serve deepseek-ai/DeepSeek-V4-Pro-0813 \
  --trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
  --tensor-parallel-size 8 --enable-expert-parallel \
  --attention_config.indexer_kv_dtype=mxfp4 --moe-backend deep_gemm_mega_moe \
  --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE", "custom_ops":["all"]}' \
  --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"probabilistic"}' \
  --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4 \
  --host 127.0.0.1 --port 8000 --api-key "$KEY"

For the most throughput with many users, the recipe recommends data parallelism instead, --data-parallel-size 8 in place of --tensor-parallel-size 8, as in DeepSeek’s own command for a 4-GPU node. For Flash, start from vLLM’s Flash recipe for your GPU and GPU count. With SGLang, here is the command from DeepSeek’s model card, set to eight GPUs and bound to 127.0.0.1, which is also SGLang’s default host (server arguments):

sglang serve --trust-remote-code \
  --model-path deepseek-ai/DeepSeek-V4-Pro-0813 --tp 8 \
  --moe-runner-backend flashinfer_mxfp4 --speculative-algorithm DSPARK \
  --mem-fraction-static 0.90 --chunked-prefill-size 4096 --swa-full-tokens-ratio 0.1 \
  --host 127.0.0.1 --port 30000 --api-key "$KEY"

# on your computer: the vLLM API is then at http://localhost:8000/v1
ssh -N -L 8000:127.0.0.1:8000 root@203.0.113.10

Replace 203.0.113.10 with your server’s address, and forward port 30000 for SGLang. With vLLM, clients pick the reasoning effort per request, for example "chat_template_kwargs": {"thinking": true, "reasoning_effort": "high"}; the max effort needs a --max-model-len of at least 393,216 tokens. Installation, Docker and a TLS proxy are covered in serving LLMs with vLLM.

What it costs to self-host DeepSeek-V4

A monthly server has one price, however many tokens it generates. The cheapest fits at GPUCenter:

• Against an hourly GPU cloud. Our prices are about 30% below the median public on-demand price of 40 GPU clouds: eight B200 at that median cost $39,360 for a month around the clock. Against that median, a monthly server pays off once it runs more than about 510 hours a month (break-even for each GPU); for a few days of tests, hourly billing costs less.
• Against a per-token API. Divide the monthly price by the API’s price per million tokens to find the volume at which the server pays off, then check that the server keeps up with that volume. DeepSeek’s API no longer serves V4-Flash, so for that exact model the choice is your own server or a third-party host.
• Against buying the hardware. An 8-GPU B200 or B300 server is a large up-front purchase, plus power, cooling and hosting. Renting turns it into a monthly cost that you can stop.

Caveats before you rent

• Long contexts and many users need more memory. Our 20% margin covers moderate contexts. V4’s KV cache is small for its context length, but a 1M-token window, max reasoning effort and many parallel conversations add up: that is why vLLM caps Pro at 800,000 tokens on 8× H200. Watch the KV cache usage once the model runs.
• Every expert stays in memory. Flash computes each token with 13B parameters, but all 284B must be loaded in GPU memory, and all 1.6T for Pro.
• Verified setups vary. vLLM’s recipe has no verified H100 profile for Flash yet, and our catalog does not confirm ROCm support for V4. Test the exact engine version, GPU and flags you plan to run.
• Plan the download. The checkpoints weigh 167 GB and 893 GB. Each B200 or B300 comes with 3.84 TB of local NVMe, 30.72 TB in an 8-GPU server.
• No benchmarks from us. We publish no speed measurements. Rent a server for a month and measure tokens per second on your own prompts and concurrency.

Where GPUCenter fits

GPUCenter rents dedicated GPU servers by the month: up to 8 GPUs per server, root SSH access and no other tenant on your cards. At the Software step of the deploy page, we can preinstall DeepSeek-V4-Flash or DeepSeek-V4-Pro with vLLM or SGLang at no extra cost, on the cheapest server that holds it, and the server is online in under 10 minutes once the payment is confirmed on-chain. You pay in Bitcoin, Ethereum, USDT, Monero or Litecoin, with no KYC.

It is the wrong choice for a few hours of tests, where an hourly cloud costs less, for a model that needs several servers joined by InfiniBand, which we do not offer, or for a managed API. We do not offer V4 on the MI355X until its ROCm support is confirmed in our catalog. For a DeepSeek model on that GPU today, our catalog confirms DeepSeek-R1-0528 on ROCm: 4× MI355X, $5,636 a month in FP8 (MI355X vs B200). To compare GPUs for serving, read the best GPU for LLM inference.

DeepSeek-V4 questions

Can DeepSeek-V4-Flash run on a single GPU?

Yes, on one B300 (288 GB), by our sizing rule: Flash needs about 200 GB. One B200 (180 GB) or H200 (141 GB) is too small; two of either hold it.

How much VRAM does DeepSeek-V4-Pro need?

About 1,072 GB by our rule, for 893 GB of official weights; vLLM’s recipe sets a 960 GB minimum. Only 8× B200 and 8× B300 servers hold it with our margin.

Can I run DeepSeek-V4 on H100 GPUs?

By our rule, Flash needs 4× H100, $7,116 a month, but vLLM’s recipe has no verified H100 profile for it yet. Pro does not fit: eight H100 have 640 GB in total.

Does DeepSeek-V4 run on the AMD MI355X?

ROCm support for DeepSeek-V4 is not confirmed in our catalog yet, so we offer it on NVIDIA servers. DeepSeek-R1-0528 is confirmed on ROCm and runs on 4× MI355X, $5,636 a month.

Can I use DeepSeek-V4 commercially?

The code and the weights are under the MIT license, which grants the right “to deal in the Software without restriction” (license), provided you keep its copyright and permission notice. You are responsible for complying with it.

What about DeepSeek-V4.1-Flash?

DeepSeek released it on 10 September 2026: a multimodal mixture-of-experts model with 552B parameters, 8B active for input and 16B for output (DeepSeek). This article covers V4-Flash and V4-Pro, the V4 models our deploy page can preinstall at the time of writing. vLLM’s recipe sets a 614 GB minimum for V4.1-Flash and requires vLLM 0.31.0 or later.