Run DeepSeek V4 on Intel® CPUs and GPUs

2026-05-18 · Source: Artificial Intelligence (AI) articles · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Intermediate, short

Summary

The recently released DeepSeek V4 introduces several frontier architectural optimizations, including a hybrid attention mechanism (Compressed Sparse Attention and Heavily Compressed Attention) that reduces KV cache usage by up to 90%. Its architecture also implements manifold-constrained Hyper-Connections (mHC) for enhanced expressiveness and training stability, and a massive Mixture-of-Experts (MoE) architecture natively trained with the MXFP4 data format, enabling advanced capabilities with a minimal computational footprint. This blog post details the steps for running DeepSeek V4 on Intel® Xeon® CPUs and Intel® Arc™ GPUs using SGLang, providing Docker-based setup and command-line instructions for both platforms to launch an OpenAI-compatible server and query models like DeepSeek-V4-Pro and DeepSeek-V4-Flash.

Key takeaway

For MLOps Engineers deploying DeepSeek V4, this guide confirms that Intel Xeon CPUs and Arc GPUs are now viable platforms. You can utilize SGLang's tailored kernels and Docker setup to achieve efficient inference, reducing KV cache usage by up to 90% and benefiting from MXFP4 MoE. Consider integrating these Intel-optimized solutions to expand your hardware options for DeepSeek V4 deployments.

Key insights

DeepSeek V4 leverages hybrid attention, mHC, and MXFP4 MoE for efficiency and expressiveness, now runnable on Intel CPUs/GPUs via SGLang.

Principles

• Hybrid attention reduces KV cache usage by up to 90%.
• mHC improves model expressiveness and training stability.
• MXFP4-trained MoE enhances computational efficiency.

Method

The article details a Docker-based setup for SGLang, followed by launching an OpenAI-compatible server for DeepSeek V4 models on Intel Xeon CPUs or Arc GPUs, then querying via curl.

In practice

• Use SGLang's Dockerfiles for Intel CPU/GPU environment setup.
• Launch an OpenAI-compatible server with `sglang serve`.
• Query DeepSeek V4 models via standard API calls.

Topics

• DeepSeek V4
• Intel Xeon CPUs
• Intel Arc GPUs
• SGLang
• Mixture-of-Experts
• LLM Inference
• Sparse Attention

Code references

• sgl-project/sglang
• sgl-project/sglang

Best for: Machine Learning Engineer, AI Engineer, MLOps Engineer

Related on AIssential

• DeepSeek v4: the most expected open-source model ever released, and the quietest landing — DeepSeek v4 prioritizes serving cost reduction through architectural innovation over raw benchmark performance.
• 2025 Interconnects year in review — DeepSeek's open-weight, efficient MoE and MLA architectures challenge AI's cost curve, intensifying the global AI race and geopolitical tech competition.
• DeepSeek resurfaces with cheap, capable V4 — DeepSeek V4 offers competitive AI model performance at lower costs, supported by Huawei chips.
• DP Attention and TBO for DeepSeek-V4 on MI355X — ATOM optimizes MoE inference by coordinating DP Attention and overlapping standard collectives with compute.
• ChinAI #356: DeepSeek as Road Builder [修路人] — DeepSeek V4 achieves near-frontier AI capabilities with significant compute efficiency, particularly for long-context inference.
• Introducing DeepSeek V4 Flash and V4 Pro in Microsoft Foundry — Optimal AI system design balances model quality, speed, and cost by orchestrating multiple specialized models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence (AI) articles.