DeepSeek-V4-Pro hardware requirements: eight H200 NVL or a DGX B300

Which server DeepSeek-V4-Pro needs

DeepSeek-V4-Pro needs either a server with eight H200 NVL or a DGX B300. Its official -0813 release weighs 892.7 GB, 831.4 GiB, which eight H200 NVL hold with 155.5 GiB of cache to spare by our users-per-card rule, enough for about 65 concurrent conversations of 32K tokens with tensor parallelism over all eight cards. A server with eight RTX PRO 6000 Server Edition cards misses the weights by 167.1 GiB, and a DGX B300 holds them with several hundred conversations of room while computing the FP4 experts natively.

The two systems differ a great deal. The eight-card H200 NVL server is built to order; it has two NVLink domains of four cards joined over PCIe, and Hopper does not compute the FP4 weights natively. The DGX B300, which we quote per project, joins all eight GPUs over NVLink, computes FP4 natively and has more than twice the memory. Our DeepSeek hardware guide covers V4-Flash, V4.1-Flash and V3.2; this page covers the 1.6T flagship.

Our estimates by the users-per-card rule for the 831.4 GiB checkpoint, 16-bit cache counted in full on every card. DGX B300 at about 270 GB per GPU (MLPerf® accelerator name B300-SXM-270GB), read cautiously as 10⁹ bytes. Power of the PCIe servers is our wall estimate; DGX B300 from NVIDIA's user guide and datasheet as quoted in our DGX B300 articles.

DeepSeek-V4-Pro-0813 on Hugging Face

DeepSeek's card calls the -0813 checkpoint "the official release of DeepSeek-V4-Pro, superseding the preview version". vLLM's recipe, updated on 13 October 2026, describes it as "a 1.6T-total / 49B-active Mixture-of-Experts model" and makes -0813 its default. We summed the safetensors files on the repository's file list on 24 October 2026.

DeepSeek-V4-Pro-0813 model card, file list and config.json on Hugging Face, read on 24 October 2026; parameters from vLLM's DeepSeek-V4-Pro recipe (updated 13 October 2026), where the card's metadata field shows 1.7T. Checkpoint size is our sum of the listed files.

The recipe gives "~893 GB on disk versus ~865 GB for the preview", the difference being the DSpark module for speculative decoding that the release carries in the same checkpoint. The card recommends "a maximum output length of 384K tokens" for its high and max reasoning levels, so the context you plan for decides more than the 32K figures suggest. DeepSeek's own example serves the model "on a single 4×GB300 node", a rack-scale system outside what we supply.

KV cache: compressed attention, copied on every card

V4 stores far less cache per token than DeepSeek-V3.2. The config.json lists compression ratios in 62 entries (the model has 61 layers plus one multi-token-prediction layer): 29 entries compress by 4, 30 by 128 and three are 0, which by our reading keeps only the 128-token sliding window. We count the 512-value compressed entry and the 128-value indexer key in 16-bit on the layers that compress by 4, and the entry alone on those that compress by 128, the same count that gives 6.4 KiB per token for V4-Flash in our DeepSeek guide. For V4-Pro that is about 9.3 KiB per token.

Our estimate from the compress_ratios, head_dim and index_head_dim fields of config.json, read on 24 October 2026; the sliding window adds about 7.8 MiB per conversation. An FP8 cache, which vLLM's recipes set with --kv-cache-dtype fp8, is smaller.

The order of magnitude matches the recipe, which credits V4-Pro with "10% of V3.2's KV cache at 1M context"; our count for V3.2 at 32K is 2.39 GiB. The catch is where the cache lives. vLLM's blog post of 7 August 2026 states for latent caches: "Under normal TP there is nothing to split by head, meaning the latent KV cache is replicated in full across every TP rank." V4 has one KV head, so each card in a tensor-parallel layout holds every conversation's whole cache, and the spare memory of one card, not of the server, sets the number of conversations.

Eight H200 NVL: the weights fit with FP4 experts in 4-bit

Eight H200 NVL give 986.9 GiB by our rule. After 831.4 GiB of weights, 155.5 GiB remain, 19.4 GiB per card. Hopper has FP8 but no FP4 arithmetic, as our guide to FP4 models on the H200 NVL and Ada GPUs explains. The recipe still lists "H200 (8× GPU): DP + EP with --data-parallel-size 8" and caps the context at 800K tokens "to leave KV headroom with dense params replicated across ranks". A copy of 1.6T parameters in FP8 would take about 1.6 TB, more than the 1,128 GB of eight cards, so by our arithmetic the H200 layout works only because the experts stay in 4-bit in memory and are computed at higher precision. The recipe does not say how Hopper handles them, so treat the speed on this card as something to measure.

Our estimates by the users-per-card rule, eight H200 NVL, 16-bit cache, 19.4 GiB of spare memory per card; with TP=4 × PP=2, one stage per NVLink domain, each card holds the cache of its own 31 or 30 of the 61 layers, counted for the larger stage. Upper bounds: on eight H200, vLLM's Kimi-K2-Thinking recipe measured about 7.4 GiB per card less cache than our rule, as our Kimi K2 guide describes.

The recipe's data-parallel layout gives each card its own cache but repeats the dense weights on every card, and the 800K cap shows how little room that leaves. We did not size it because the size of the dense part is not published. Two further points apply to the H200 NVL in particular. The recipe does not state the form factor of its H200; in an eight-card H200 NVL server two 4-way bridges form two NVLink domains of four, joined over PCIe, so tensor parallelism over all eight and the expert traffic of DP + EP cross PCIe, as our guide to the 8× H200 NVL server shows. TP=4 × PP=2 keeps each stage's all-reduces on NVLink and roughly halves the cache per card. We build this server with its NVLink bridges and size the layout per model; send us the context length and the peak number of conversations, and our team replies with the figures per layout.

Why eight RTX PRO 6000 Server Edition are not enough

The RTX PRO 6000 Server Edition computes FP4 natively, which suits the V4 experts, but eight cards have 768 GB, less than the 892.7 GB of the checkpoint before any cache. By our rule they hold 664.3 GiB, 167.1 GiB short of the weights. NVIDIA's variant nvidia/DeepSeek-V4-Pro-NVFP4, which vLLM's recipe lists for Blackwell, does not help either: it re-quantises the experts to standard NVFP4 and keeps attention, shared experts and the prediction module in FP8, and its 64 files sum to 913.1 GB (850.4 GiB). The smaller members of the family fit this server: DeepSeek-V4-Flash at 167 GB and DeepSeek-V4.1-Flash at about 511 GB, which our DeepSeek hardware guide sizes. For a model of 600 to 700 GB on Blackwell PCIe cards, our guide to eight RTX PRO 6000 Server Edition for LLMs lists what fits.

DGX B300: native FP4 and NVLink across eight GPUs

A DGX B300 has eight Blackwell Ultra GPUs with 288 GB of HBM3e each, of which about 270 GB is exposed, the figure in the accelerator name B300-SXM-270GB of MLPerf® entries. Read cautiously as 10⁹ bytes, that is 251.5 GiB per GPU and 223.3 GiB by our rule, 1,786.5 GiB for the system. After the weights, about 955 GiB remain, 119.4 GiB per GPU. With the cache copied to every GPU under TP=8, that is room for about 400 conversations of 32K or 34 at 384K; the recipe's B300 layout is "single-node DP + EP with --data-parallel-size 8", which we have not sized for the reason given above.

The B300 computes the FP4 experts in FP4, so no weight-only path is involved, and NVLink joins all eight GPUs, so no traffic crosses PCIe. Our comparison of DGX B300 and a server with eight RTX PRO 6000 covers the bandwidth and the MLPerf® results. The system is different in the room as well: 10U and 168 kg in the AC version, twelve 3.2 kW power supplies and 14.5 to 15.1 kW in NVIDIA's user guide and datasheet. We supply the DGX B300 per project; before we can quote, the distribution asks for the company, the project and the intended use, as our guide to buying a DGX B300 in the EU describes.

Server, power and room for each option

The eight-card H200 NVL server follows our training node: two AMD EPYC processors, 2.3 TB of DDR5 ECC, eight NVMe drives and 400G networking. The memory follows NVIDIA's configuration guide for certified systems, which asks for at least twice the GPU memory, and our guide to RAM for a GPU server explains the rule. The checkpoint needs 893 GB on local NVMe, and room for a second copy makes an update possible without taking the service down.

With eight cards at 600 W and two 64-core processors, we estimate about 6.3 kW at the wall, a figure that a measurement on the delivered system confirms or corrects. That needs a 400 V three-phase position on each of the A and B feeds, as our guide to GPU server power on 16 A, 32 A and three-phase shows. A DGX B300 draws more than twice as much: NVIDIA's data-centre guide for the system plans two AC systems per rack at 39.4 kW peak and recommends active rear-door heat exchangers for high-density racks. Check what your room or colocation contract delivers per rack before choosing. We check rack, feed and airflow before we quote; tell us the power and cooling available at the rack position, and we say which option fits.

When two H200 NVL servers are the better step

One eight-card H200 NVL server reaches its limit when many users work with long contexts at once: at 384K, the reasoning length the card recommends, it holds about 10 conversations even in the two-stage layout. Two servers with eight H200 NVL each have 1,973.8 GiB by our rule. Run as one copy with pipeline stages across both, they leave 71.4 GiB per card for cache, almost four times as much as one server. Run as two copies, they keep the service up if one server fails. Our comparison of 1T models for 100 users on a DGX B300 or two H200 NVL servers works through both layouts for V4-Pro and other large models.

MIT licence and data handling

The card states: "This repository and the model weights are licensed under the MIT License." The licence permits use, modification and distribution on condition that the copyright and permission notice are included in copies, and it excludes any warranty. A model run from downloaded weights on your own servers sends no prompts or outputs to DeepSeek. Whether further terms apply to your use is for your legal department to assess.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. For DeepSeek-V4-Pro we build servers with eight H200 NVL and two 4-way NVLink bridges, two AMD EPYC processors, 2.3 TB of DDR5 ECC, eight NVMe drives and 400G networking, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. The ordering steps are on our page to buy an 8× H200 NVL server, and the DGX B300 we quote per project. Deploying the model with vLLM and connecting it to your applications is part of our Private AI/ML service.

FAQ