DeepSeek V4 Launches With NVIDIA Blackwell, Enabling 1M-Token Context AI

DeepSeek has unveiled its fourth-generation AI models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, pushing the boundaries of long-context inference. These models, available now via NVIDIA’s Blackwell GPU-accelerated endpoints, are designed to handle up to a 1 million-token context window, a significant step forward for applications like advanced coding, document analysis, and agentic AI workflows.

The flagship DeepSeek-V4-Pro boasts 1.6 trillion total parameters with 49 billion active parameters, while the more efficiency-focused DeepSeek-V4-Flash features 284 billion total parameters and 13 billion active parameters. Both models are licensed under MIT and cater to distinct use cases—Pro for advanced reasoning and Flash for high-speed tasks like summarization and routing.

Architectural Breakthroughs for Long-Context AI

DeepSeek V4 builds on the company’s Mixture-of-Experts (MoE) architecture, introducing innovations aimed at overcoming the challenges of long-context inference. The new hybrid attention mechanism blends Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), enabling a 73% reduction in per-token inference FLOPs and a 90% reduction in KV cache memory usage compared to its predecessor, DeepSeek V3.2.

Why does this matter? As context windows expand, managing memory and compute efficiency becomes crucial. Long-context AI applications like multi-turn reasoning, tool integration, and extensive workflows require models that can retain and process large amounts of contextual data without bottlenecks. DeepSeek V4’s improvements address these pain points, making it a strong contender for enterprises aiming to scale AI-driven systems.

NVIDIA Blackwell Integration

DeepSeek V4 is tightly integrated with NVIDIA’s Blackwell platform, leveraging its GPU-accelerated infrastructure for scalable performance. Initial tests on the NVIDIA GB200 NVL72 hardware show DeepSeek-V4-Pro achieving over 150 tokens per second per user, with ongoing optimizations expected to further improve throughput.

Blackwell’s architecture is designed for trillion-parameter intelligence models, making it a natural fit for DeepSeek V4’s computational demands. Developers can prototype with these models via NVIDIA’s hosted endpoints on build.nvidia.com or deploy them directly using NVIDIA NIM for custom infrastructure setups.

Target Use Cases and Deployment Flexibility

DeepSeek V4’s ability to handle 1M-token contexts opens new opportunities for long-context coding, retrieval-based workflows, and agentic AI. Its flexibility is further enhanced by deployment tools like SGLang and vLLM, which offer recipes tailored for different latency and throughput needs, from low-latency setups to multi-GPU configurations for large-scale operations.

This focus on deployment flexibility underscores a broader trend: as open AI models approach the frontier of intelligence, enterprises are shifting their attention from model selection to infrastructure optimization. The ultimate goal is reducing the cost per token while maintaining performance, and DeepSeek V4 aligns squarely with this priority.

Getting Started

Developers can access DeepSeek V4 through multiple channels, including Hugging Face and NVIDIA’s API endpoints. For enterprises and developers looking to integrate long-context AI into their workflows, DeepSeek V4 offers a compelling combination of scalability, efficiency, and advanced reasoning capabilities.

With its architectural advancements and seamless integration with NVIDIA Blackwell, DeepSeek V4 sets a new benchmark for long-context AI. As the demand for agentic systems and expansive context windows grows, models like these will play a pivotal role in shaping the next generation of AI applications.

Image source: Shutterstock