DeepSeek V4 GPU requirements: Flash, Pro and what it takes to self-host

DeepSeek V4-Flash needs about 155 GiB of GPU memory for its weights and V4-Pro about 831 GiB, so Flash is a job for two to four large GPUs and Pro for one 8-GPU H200, B200 or B300 server. No GPU that QuantaCloud rents by the hour holds either one alone. Both are mixture-of-experts models under the MIT licence, with expert weights already stored in 4 bits, so the files are much smaller than the parameter counts suggest: 166.89 GB for V4-Flash-0731 and 892.74 GB for V4-Pro-0813.

The DeepSeek V4 models#

DeepSeek has published V4 weights in several rounds since April 2026. The two that matter today are the official releases, V4-Flash-0731 and V4-Pro-0813, which replaced the April previews:

DeepSeek's Hugging Face page also holds V4-Flash-Vision-Exp, an experimental multimodal Flash at 167.82 GB. The official releases are 7.3 GB and 28.0 GB larger than the previews because DeepSeek attached a DSpark speculative-decoding module to each (our calculation from the file sizes).

DeepSeek calls the format "FP4 + FP8 Mixed": the experts are stored in FP4 and most other parameters in FP8. Flash has 43 layers with 256 routed experts each and Pro has 61 layers with 384, and both use 6 experts per token. The Flash, Pro and V4.1 model cards carry the same sentence: "This repository and the model weights are licensed under the MIT License." None of the repositories is gated.

One thing to know before you plan around V4: DeepSeek's own API has moved on. Its 2026-09-10 announcement retired V4-Flash and said that from 2026-09-14 all deepseek-v4-pro requests route to V4.1-Flash until a V4.1-Pro launches. The V4 weights stay on Hugging Face under the same licence, so you can still run them yourself.

Why memory follows the total parameters#

A mixture-of-experts model reads only a few experts for each token, but which ones changes token by token, so every expert has to sit in GPU memory. Size V4 on its total parameters and its file size, never on the 13B or 49B active count. The VRAM guide covers the general rule.

The 4-bit experts are what make V4 practical, and they are not optional. DeepSeek also publishes the pre-trained Base models with FP8 experts, and they show what the larger format costs: 294.68 GB for V4-Flash-Base against 159.62 GB for the preview Flash, and 1,606.05 GB for V4-Pro-Base. Serving the FP4 files needs FP4 kernels in the engine, which is why the list of supported GPUs is short.

The KV cache is the small part. V4's compressed attention stores far less per token than earlier DeepSeek models: the model card puts V4-Pro at 10% of V3.2's cache at a million tokens. DeepSeek's V4.1 card gives 890 bytes per token for the global cache of V4.1-Flash and calls that roughly a quarter of V4-Flash, which makes V4-Flash about 3.6 KB per token and a full million-token conversation about 3.5 GiB (our calculation). The context you allow still matters, because the highest reasoning effort wants a window of at least 384K tokens.

Which configurations hold V4-Flash#

V4-Flash-0731 is 155.43 GiB of weights. The budget is 92% of the memory the GPUs report in nvidia-smi, vLLM's default (our calculation):

On demand, that points at two VMs: 2x H200 NVL and 4x RTX PRO 6000. Both hold the weights with room to spare, and neither is a layout vLLM has verified, so test before you depend on one. They are also not always in the catalog. Launch one when the catalog lists it, and send a capacity brief for a built-to-order server when it does not.

No configuration is listed right now. See every GPU in the console.

Some configurations fit on size and are still a poor bet. vLLM's recipe supports the 0731 checkpoint on Hopper cards, on B200 and B300, and on an 8-GPU RTX PRO 6000 server. Ampere and Ada cards are not on that list, so I would not rent a 4x A100 or an 8x L40S VM for it, even though both have the memory. Two H100 PCIe cards are 8.9 GiB short. The layout the recipe recommends is data parallelism with expert parallelism across four GPUs, which uses four of the eight GPUs in an H200, B200 or B300 server for each copy of the model.

V4-Pro needs an 8-GPU server#

V4-Pro-0813 is 831.43 GiB of weights, more than any on-demand VM offers. The largest one, 8x A100 SXM4, has a budget of 588.8 GiB. These are the servers that hold it (our calculation):

So the answer for Pro is about 831 GiB of GPU memory before any cache, which is one 8-GPU server of H200s or larger. On H100s the weights alone need twelve GPUs by size, which means two 8-GPU servers. We build servers like these to order: send a capacity brief with the model and the traffic you expect, and the configuration, lead time and terms come back in writing. H200, B200 and B300 describe the GPUs, and dedicated GPU servers describes the offer.

DeepSeek V4.1-Flash#

V4.1-Flash is the model DeepSeek now serves, and it sits between Flash and Pro in size: 510.30 GB on disk, 475.25 GiB. It is a different architecture, a multimodal model with a 552B backbone and a 196B lookup memory that DeepSeek calls Engram, and the 1017 revision changed the weights without changing the size.

vLLM's recipe says it fits one 8-GPU H200 server with room for the cache, and that an 8-GPU server of 80 GB cards is the smallest NVIDIA node that serves it, with the Engram tables moved to host memory. The recipe verifies that on H100s. The 8x A100 SXM4 VM has the same memory but is not on its list. On four H200s with that offload, the recipe reports 81.2 GiB of weights and 38.5 GiB of cache per GPU. It needs vLLM 0.31.0 or newer. For a new self-hosted DeepSeek deployment on an 8-GPU H200 server, this is the model I would compare against V4-Pro first, since DeepSeek itself replaced Pro with it.

Engine support: vLLM, GGUF and Ollama#

vLLM is the engine DeepSeek documents, along with SGLang. The recipe gives vLLM 0.25.0 as the minimum for the 0731 and 0813 checkpoints, the release that added DSpark drafting, and pins the v0.28.0 image for V4-Flash-0731. The V4 architecture is registered in vLLM 0.30.0 and in 0.33.0, the latest release on 2026-11-12. This is DeepSeek's own example from the V4-Flash-0731 model card, written for a four-GPU GB300 node:

vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \
  --trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
  --data-parallel-size 4 --enable-expert-parallel \
  --moe-backend deep_gemm_mega_moe \
  --attention-config '{"use_fp4_indexer_cache": true}' \
  --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

The MoE backend and the FP4 indexer cache in that command are kernels for data center Blackwell GPUs such as the B200 and GB300. vLLM's recipe drops both on the H200 and on the RTX PRO 6000, a different Blackwell chip, so build the command for your GPU from the recipe instead of copying this one. The V4 repositories ship no Jinja chat template, only Python encoding scripts. In vLLM, --tokenizer-mode deepseek_v4 applies the same encoding inside the server, so the OpenAI-compatible chat endpoint works without those scripts. The vLLM multi-GPU guide explains tensor, data and expert parallelism.

GGUF files exist, but not from DeepSeek. Unsloth publishes community quantizations of V4-Flash-0731 from 82.54 GB to 161.87 GB, and llama.cpp has supported the V4 architecture since June 2026. The saving is smaller than usual, because the original is already 4-bit: the Q4 file is 155.10 GB against 166.89 GB. The files that are meaningfully smaller are 2-bit and below, such as the 96.83 GB one, which fits a single H200 NVL on size, and they give up quality to get there. I would use them to try the model, not to judge it. The llama.cpp guide covers splitting a GGUF across GPUs.

Ollama does not offer V4 as a download. Its deepseek-v4-flash page lists no models, and deepseek-v4-pro and deepseek-v4.1-flash list only cloud tags, which run on Ollama's servers and never reach your GPU.

DeepSeek V4 FAQ#

Can I run DeepSeek V4 locally?

Not on one GPU, unless it is a B300 or a GB300. V4-Flash needs about 155 GiB for its weights, which is two H200s or four RTX PRO 6000s by size, and V4-Pro needs an 8-GPU server. The desk-side machines that run V4-Flash are NVIDIA's DGX Station, whose single GB300 GPU vLLM's recipe has verified, and a pair of DGX Spark units with a community vLLM build.

How much VRAM does DeepSeek V4-Flash need?

155.43 GiB for the weights of V4-Flash-0731, plus a few GiB of cache per million tokens of context. Plan for a 2x H200 NVL or 4x RTX PRO 6000 configuration as the minimum.

Is there a DeepSeek V4 GGUF?

Yes, from the community. Unsloth's V4-Flash-0731 files run from 82.54 GB to 161.87 GB, and its V4-Pro-0813 files are 849.68 GB and 873.45 GB. DeepSeek publishes safetensors only.

Is DeepSeek V4 on Ollama?

Only as cloud tags for V4-Pro and V4.1-Flash, which Ollama hosts. There is no V4 tag that downloads.

Is DeepSeek V4 free for commercial use?

Yes. The V4 weights are under the MIT licence.

The rule I follow for DeepSeek V4 is to size it by the file and the GPU family together: Flash on two H200s or four RTX PRO 6000s when the catalog lists them, and Pro or V4.1-Flash on an 8-GPU H200, B200 or B300 server built to order. For the older models, the R1 distills that fit one GPU and the 671B R1 and V3 weights, the DeepSeek GPU requirements guide has the numbers. Self-hosting an LLM covers the decision around the hardware.