DeepSeek V4 on NVIDIA: Specs, GPU Needs & Cost

Last updated: June 2026 · By Ignacy Kwiecień, founder & editor-in-chief, DecodeTheFuture.org

What are DeepSeek V4-Pro and V4-Flash specs, side by side?

DeepSeek V4 is a two-model family, not one model, and that distinction is the single thing most spec recaps blur. V4-Pro is the flagship Mixture-of-Experts model at roughly 1.6 trillion total parameters with about 49B activated per token — only the selected experts fire on any given token, so it carries frontier-scale knowledge at a fraction of a dense 1.6T model’s per-step compute. V4-Flash is the lighter sibling at 284B total / 13B activated, built for cheaper, faster serving where you do not need Pro’s full reasoning depth.

Both members share a 1 million-token context window and ship under the permissive MIT License, which is unusually liberal for a model of this size — you can use it commercially without the bespoke community-license clauses some competitors attach. V4-Pro’s headline architectural move is a hybrid attention stack: Compressed Sparse Attention (CSA) plus Heavily Compressed Attention (HCA), the mechanism behind the ~90% cut in KV-cache memory versus V3.2 in the 1M-token context setting that DeepSeek reports on its model card.

One spec deserves a flag: the V4-Flash 284B / 13B figure is vendor-listed and we could not independently verify it against a stable primary model card at the time of writing, so treat it as directional rather than a confirmed audited number. The V4-Pro 1.6T / 49B figure is confirmed in NVIDIA’s own Dynamo release notes and is the safer one to plan around.

What reasoning modes does DeepSeek V4 have, and why do they matter for cost?

DeepSeek V4 exposes three reasoning modes — Non-Think, Think High and Think Max — and they are the most direct lever you have over your inference bill. Non-Think answers without an extended reasoning trace, which is cheapest and fastest. Think High allocates a moderate reasoning budget. Think Max unlocks the deepest chain and is where V4-Pro posts its strongest benchmark numbers.

Those benchmarks are V4-Pro’s selling point, and they come from the model card in Max reasoning mode: 80.6 Resolved on SWE-bench Verified (a real-world coding-agent benchmark), 93.5 Pass@1 on LiveCodeBench, and 87.5 EM on MMLU-Pro. The practical takeaway for a buyer: you only pay for Think Max’s token cost when you actually need that depth. For high-volume agent loops, defaulting to Non-Think and escalating to Think Max only on hard turns is the same cost discipline you would apply to retries — budget reasoning per call, not globally.

What GPUs and VRAM does DeepSeek V4 need on NVIDIA Blackwell?

There is no consumer-GPU or single-card path to V4-Pro — it is a multi-GPU, datacenter-class deployment. The hardest number to pin down across the SERP is VRAM, because sources quote wildly different figures depending on precision. A widely cited third-party estimate puts V4-Pro’s weights near ~500GB at FP8 and ~1TB at BF16 (Spheron’s deployment guide — treat as a directional estimate, not an NVIDIA-audited figure). Either way you are well past a single 80GB card, which is why NVIDIA’s own reference points are whole nodes, not chips.

The serving format matters as much as raw weight size, and it helps to separate two things. DeepSeek’s model card describes the model’s own stored weight precision as FP4 + FP8 mixed (MoE experts in FP4, the remaining parameters in FP8). NVIDIA then serves it in its native MXFP4 format on Blackwell, further tuned with NVFP4 and custom CUDA kernels — the OCP MXFP4 serving format is NVIDIA’s runtime representation, not identical to DeepSeek’s storage precision. Low-precision formats shrink both the weight and activation footprint, which is what lets the model fit fewer GPUs and stretch toward its 1M context. The deeper mechanics of why low-precision KV cache buys you context headroom are their own subject — if you want that, see our explainer on vector quantization and the KV cache rather than re-deriving it here.

For a concrete NVIDIA-confirmed reference, the Dynamo dev build that serves V4-Pro ships an example configuration on 4× GB300. That is a Blackwell-generation node, and it is the closest thing to an official “this is what it runs on” number currently published. The FP8/BF16 footprint figures above come from third-party guides, so use them to sanity-check your own hardware, not to size a final budget.

How do you run DeepSeek V4 with Dynamo and TensorRT-LLM?

The shipped path to self-serve V4-Pro on NVIDIA is the Dynamo dev build v1.3.0-deepseek-v4-dev.1, released June 6, 2026, with its TensorRT-LLM backend pinned to 1.3.0rc15.post1. That pin is not optional flavour text — the V4-Pro attention kernels (CSA/HCA) need that exact TensorRT-LLM build to serve correctly, which is the gotcha most third-party guides skip. The same June 2026 Dynamo wave shipped day-0 dev builds for MiniMax-M3, Kimi K2.6 and Nemotron-3 Super alongside DeepSeek-V4-Pro.

# Dynamo dev build for DeepSeek-V4-Pro (June 6, 2026)
# Backend is pinned to TensorRT-LLM 1.3.0rc15.post1 - the
# CSA/HCA attention kernels need this exact build to serve.
# NVIDIA's example config targets 4x GB300 (Blackwell).
#
# WARNING: this is an EXPERIMENTAL dev build, NOT for production.

git clone https://github.com/ai-dynamo/dynamo.git
cd dynamo
git checkout v1.3.0-deepseek-v4-dev.1

# TensorRT-LLM backend pin (required):
#   tensorrt-llm==1.3.0rc15.post1

# Launch the Dynamo serving graph on a 4x GB300 node
# (see release notes for the deepseek-v4-pro example config)
dynamo serve graphs.deepseek_v4_pro:Frontend \
  --tensor-parallel-size 4 \
  --enable-think-modes  # Non-Think / Think High / Think Max

If you do not need to stand up your own serving graph, TensorRT-LLM itself carries the V4 attention kernels and AutoConfig support in its recent releases, so it is the lower-level building block under both Dynamo and the NIM container. The runtime landscape (vLLM, TensorRT-LLM, SGLang) is a comparison in its own right and we do not benchmark them against each other here — for V4-Pro specifically, the NVIDIA-blessed path is TensorRT-LLM, because that is where the CSA/HCA kernels live.

How do you access DeepSeek V4 day-0 as an NVIDIA NIM?

DeepSeek V4 is available day-0 as an NVIDIA NIM container and through GPU-accelerated endpoints on build.nvidia.com, which is the fastest way to call the model without owning a Blackwell node. The NIM packages V4 with NVIDIA’s optimized serving stack — the MXFP4/NVFP4 kernels, TensorRT-LLM backend and the reasoning-mode controls — behind an OpenAI-compatible API. If NIM is new to you, our NVIDIA NIM API explained primer covers how the microservice works internally, and you can grab a key via the free NVIDIA API key walkthrough before you call it.

# Day-0 hosted endpoint on build.nvidia.com (OpenAI-compatible).
# No GPUs to provision - pay per token, escalate reasoning per call.
curl https://integrate.api.nvidia.com/v1/chat/completions \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-ai/deepseek-v4-pro",
    "messages": [{"role":"user","content":"Plan a 3-step refactor."}],
    "reasoning_mode": "think_high"
  }'

The hosted route is where most teams should start. It sidesteps the FP8 ~500GB footprint, the 4× GB300 floor and the experimental-dev-build caveat entirely — you validate V4’s quality on your own prompts first, measure real tokens-per-turn, and only then decide whether a dedicated node is worth it. For a broader view of how NVIDIA’s hosted offering stacks up against managed substitutes, our best inference APIs 2026 roundup and the NVIDIA NIM alternatives guide map the wider market.

When does self-hosting DeepSeek V4 beat the hosted endpoint?

The hosted NIM endpoint wins until your V4 token volume is high and sustained enough to keep a 4-plus-GPU Blackwell node near full utilization 24/7. Below that line you are paying for idle GB300s; above it, a busy node amortizes the hardware and the per-token economics flip in your favour. NVIDIA frames the economics with a per-user throughput figure — over 150 tokens/sec/user on GB200 NVL72 — and a 30× better perf-per-watt than H200 on Blackwell (both vendor-reported). Those numbers describe a busy, well-utilized node, which is exactly the condition under which self-hosting makes sense.

The serving stack underneath this matured fast: NVIDIA Dynamo 1.0 entered production on March 16, 2026 and, per NVIDIA, boosts Blackwell inference up to 7×. So the V4-Pro dev image sits on top of a serving layer that already has a stable 1.0 production lineage — the model integration is experimental, the underlying Dynamo platform is not. The diagram makes the decision boundary concrete.

Concretely: a developer evaluating V4 on a handful of prompts, or a product with bursty traffic, has no business standing up four GB300s and pinning a release-candidate TensorRT-LLM build. The day-0 NIM absorbs the spikes and you pay only for what you use. Self-hosting earns its place when a fleet of agents runs continuously, when data residency forbids an external API, or when sustained latency control justifies dedicated hardware — and even then, V4-Pro’s serving image is still flagged experimental, so a production self-host means accepting that caveat. For the wider hosted-vs-managed cost picture, the inference APIs roundup works through the same break-even maths.

What are the caveats before you commit to DeepSeek V4 in production?

Three things will shape a real deployment more than the spec sheet.

1. The V4-Pro serving image is experimental. The Dynamo v1.3.0-deepseek-v4-dev.1 build is explicitly a dev build, not for production. For anything customer-facing today, the day-0 NIM or hosted endpoint is the supported path; the dev build is for kernel testing and evaluation.

2. The benchmarks and economics are vendor-reported. The 80.6 SWE-bench, 93.5 LiveCodeBench and 87.5 MMLU-Pro scores are from the DeepSeek model card in Max mode; the >150 tok/s/user and 30× perf/watt figures are NVIDIA’s own, while the ~73% fewer FLOPs and ~90% KV-cache figures come from the DeepSeek model card, measured in the 1M-token context setting. Use them as directional benchmarks to justify a pilot, not audited guarantees to size a cluster. Run your own workload first.

3. Some specs are firmer than others. V4-Pro’s ~1.6T/49B and the Dynamo/TensorRT-LLM pins are confirmed in NVIDIA’s release notes. The V4-Flash 284B/13B figure and the FP8 ~500GB footprint are vendor-listed or third-party estimates — verify them against the live model card before you bet hardware on the exact numbers.

FAQ