Technical Blog
Related Resources
Agentic AI / Generative AI

Build with DeepSeek V4 Using NVIDIA Blackwell and GPU-Accelerated Endpoints

Apr 24, 2026
By Anu Srivastava and Eduardo Alvarez
Discuss (0)

DeepSeek just launched its fourth generation of flagship models with DeepSeek-V4-Pro and DeepSeek-V4-Flash, both targeted at enabling highly efficient million-token context inference.

DeepSeek-V4-Pro is the largest model in the family, with 1.6T total parameters and 49B active parameters. DeepSeek-V4-Flash is a smaller 284B-parameter model with 13B active parameters, designed for higher-speed, higher-efficiency workloads. Both models support up to a 1M-token context window, opening new possibilities for long-context coding, document analysis, retrieval, and agentic AI workflows.

SpecificationDeepSeek-V4-ProDeepSeek-V4-Flash
ModalityTextText
Total parameters1.6T284B
Active parameters49B13B
Context length1M tokens1M tokens
Max output lengthUp to 384K tokens through DeepSeek API docsUp to 384K tokens through DeepSeek API docs
Primary use casesAdvanced reasoning, coding, long-context agentsHigh-speed efficiency, chat, routing, summarization
LicenseMITMIT
Table 1. Specifications for the DeepSeek V4 model family.

Architectural innovations for long-context inference

The V4 family builds on the DeepSeek MoE architecture, with an increased focus on optimizing the attention component of the transformer architecture. These innovations are designed to achieve a 73% reduction in per-token inference FLOPs and a 90% reduction in KV cache memory burden compared with DeepSeek-V3.2.

That matters because long context is becoming a core requirement for agentic applications. Agents store more than a single prompt and response. They carry system instructions, tool outputs, retrieved context, code, logs, memory, and multi-step reasoning traces across a workflow. As context windows grow, attention and KV cache become major bottlenecks.

A technical diagram illustrating the "Hybrid Attention" mechanism within a Transformer Block. It shows an Attention Block feeding into three layers: Compressed Sparse Attention (CSA), DeepSeek Sparse Attention (DSA), and Heavily Compressed Attention (HCA). This architecture is equated to a "Compressed KV Cache" and "Sparse Attention Matrices," highlighting benefits like reduced memory burden and computational overhead before the data moves to a Mixture-of-Experts (MoE) block.
Figure 1. Overview of the DeepSeek-V4 Hybrid Attention architecture. By combining CSA, DSA, and HCA, the model achieves a significant reduction in KV cache memory footprint and computational overhead within the standard transformer block.

The core architectural solution to this challenges is hybrid attention, which blends together: 

  • Compressed Sparse Attention (CSA): Leverages dynamic sequence compression to compress KV entries to reduce the KV cache memory footprint and then applies DeepSeek Sparse Attention (DSA) to sparsify the attention matrices and reduce computational overhead. 
  • Heavily Compressed Attention (HCA): Applies much more aggressive compression by consolidating KV entries across sets of tokens into a single compressed entry, resulting in significant reduction in KV cache size. 

DeepSeek-V4’s architectural innovations signal a shift from basic chat toward multi-turn, long-context inference and agentic systems. This new paradigm stresses the entire stack – software, memory, compute, and networking – fundamentally altering the dynamics of inference economics. As open models reach the frontier of intelligence, the enterprise focus is pivoting from model selection to infrastructure strategy. In this landscape, the ultimate competitive advantage is the ability to deploy and scale these high-performance models at the lowest token cost. 

Out-of-the-box NVIDIA Blackwell performance insights 

Whether developers are deploying the 1.6T Pro model for advanced reasoning or the 284B Flash model for high-speed efficiency, Blackwell provides the scale and low-latency performance required for a new era of 1M long-context inference and trillion-parameter intelligence.

The NVIDIA Blackwell Platform is built for this class of workload. Out of the box tests on DeepSeek-V4-Pro on NVIDIA GB200 NVL72 demonstrate over 150 tokens/sec/user. In addition to these initial tests, the NVIDIA team leveraged vLLM’s Day 0 NVIDIA Blackwell B300 recipe to produce a snapshot of out-of-the-box performance across the pareto (Figure 2).

A line graph titled "NVIDIA Blackwell Ultra DeepSeek-V4-Pro Performance," showing out-of-the-box performance for the 1M Long-Context model. The y-axis represents Throughput (TPS per GPU), and the x-axis represents Interactivity (TPS per User). The curve shows a sharp exponential decay, where higher interactivity (more responsive user experience) results in significantly lower total AI factory throughput.
Figure 2. Out-of-the-box performance of DeepSeek-V4-Pro on NVIDIA Blackwell B300. Results are based on a 1K/1K ISL/OSL configuration using the model’s native MXFP4 format.

Expect this performance to climb even higher as we optimize our entire extreme co-design stack: Dynamo, NVFP4, optimized CUDA kernels, advanced parallelization techniques, and beyond.  

Build with NVIDIA GPU-accelerated endpoints

Developers can start building with DeepSeek V4 through NVIDIA GPU-accelerated endpoints on build.nvidia.com as part of the NVIDIA Developer Program. Hosted endpoints provide a fast way to prototype with the latest models before moving to self-hosted deployment paths.

DeepSeek V4 is also available to download on day-0 with NVIDIA NIM so it can be deployed to build long-context coding, document analysis, and agentic workflows using familiar API patterns.

Deploying with SGLang

SGLang offers three primary serving recipes for DeepSeek‑V4 on NVIDIA Blackwell and Hopper, each tuned for a different latency/throughput profile (low‑latency, balanced, and max‑throughput), along with specialized recipes for long‑context workloads and for prefill/decode disaggregation.

Deploying with vLLM

vLLM provides DeepSeek‑V4 single‑node and multinode serving recipes for NVIDIA Blackwell and Hopper, including multinode prefill/decode disaggregation recipes scaling up to 100+ GPUs, with support for tool calling, reasoning, and speculative decoding.

Powering agentic workflows

DeepSeek V4 is especially great for agents as it excels at long context orchestration, reasoning, and tool calling. To get started, developers can configure DeepSeek V4 as the LLM:

The best part of using open agent harnesses and open models is you’re always able to try new models to pick up the bleeding edge.

Get started with DeepSeek

From data center deployments on NVIDIA Blackwell to managed NIM microservices and fine-tuning workflows, NVIDIA provides a range of options for integrating DeepSeek and other open models across different stages of development and deployment. NVIDIA is an active contributor to the open-source ecosystem and has released several hundred projects under open-source licenses. NVIDIA is committed to optimizing community software and open models lets users broadly share work in AI safety and resilience.

To get started, check out DeepSeek-V4 on Hugging Face or test out pro on build.nvidia.com.

Discuss (0)

Tags

Agentic AI / Generative AI | General | Blackwell | Intermediate Technical | Deep dive | featured

About the Authors

Avatar photo
About Anu Srivastava
Anu Srivastava is a senior technical marketing manager who focuses on NVIDIA’s lighthouse AI model collaborations. She works with key partners and foundations to enable NVIDIA accelerated platform support for the open source developer ecosystem. Prior to NVIDIA, she worked at Google for over a decade in various engineering and management roles and holds a degree in computer science from the University of Texas at Austin.
Avatar photo
About Eduardo Alvarez
Eduardo Alvarez is a senior technical lead at NVIDIA, where he focuses on AI inference at scale, performance optimization, workload economic analysis, and application enablement. He has a deep background in AI systems engineering, workload optimization, and accelerated computing—focused on translating innovations into real-world applications. Before NVIDIA, Eduardo held engineering roles at various semiconductor and energy tech companies.

Comments

Comments are closed.