DeepSeek V4 is DeepSeek's upcoming flagship AI model, expected ~1 trillion parameters with 1.8x faster inference than V3 and a 1M token context window. It's the first major frontier model trained and running entirely on Huawei Ascend chips — not Nvidia.

Why does DeepSeek V4 use Huawei chips instead of Nvidia?

US export controls limit Chinese access to Nvidia's top chips. DeepSeek spent Q1 2026 porting its MLA and MoE frameworks to Huawei Ascend 950PR NPUs, proving frontier-scale models can train and run on Chinese-made hardware.

When will DeepSeek V4 be released?

April 2026 release target. Chinese tech giants including Alibaba are preparing infrastructure for the launch. It marks a significant milestone in China's AI independence strategy.

Quick Answer

DeepSeek V4 on Huawei: China's AI Chip Breakthrough

Published: April 5, 2026

DeepSeek V4 on Huawei: China’s AI Chip Breakthrough

DeepSeek V4 is China’s next-generation AI model — and it will run entirely on Huawei chips, not Nvidia. This is a pivotal moment in China’s AI independence strategy.

Last verified: April 2026

Quick Facts

Detail	Info
Model	DeepSeek V4
Architecture	~1 trillion parameters (MoE)
Chips	Huawei Ascend 950PR NPUs
Context window	1M tokens
Inference speed	1.8x faster than V3
Release	April 2026 target

Why It Matters

The Nvidia Independence Milestone

Until now, all frontier AI models — GPT, Claude, Gemini, Llama — trained on Nvidia GPUs. DeepSeek V4 breaks that pattern. By running entirely on Huawei Ascend chips, DeepSeek proves Chinese hardware can handle trillion-parameter model training.

This matters because:

US export controls limit China’s access to Nvidia’s top chips (H100, H200, Blackwell)
China’s AI ambitions require domestic chip capabilities
Global AI landscape could bifurcate into Nvidia/CUDA and Huawei/CANN ecosystems

The Engineering Challenge

DeepSeek spent Q1 2026 working with Huawei engineers to:

Port Multi-head Latent Attention (MLA) to Ascend NPUs
Adapt DeepSeekMoE framework to Huawei’s CANN toolkit
Close the maturity gap between CUDA and CANN
Fix training failures that affected the earlier R2 model on Ascend 910C

Technical Specs (Expected)

Parameters: ~1 trillion (Mixture of Experts)
Context: 1M tokens
Features: Engram memory system for long-context efficiency
Training: Hundreds of thousands of Ascend 950PR chips
Inference: 1.8x faster than V3 at comparable quality

Implications

For China

Proves domestic AI infrastructure is viable at frontier scale
Huawei’s Ascend ecosystem gains credibility
Alibaba and other Chinese tech giants preparing to adopt

For the Global AI Market

Nvidia’s moat narrows slightly
Creates a parallel AI stack (Huawei/CANN) alongside Nvidia/CUDA
Potential for a bifurcated AI ecosystem (Western vs Chinese)

For Developers

DeepSeek V4 likely open-sourced (DeepSeek’s tradition)
Cheaper inference than US models (historical pattern)
Can be deployed on Huawei cloud without export restrictions

DeepSeek V4 vs Competition

Model	Parameters	Context	Chips
DeepSeek V4	~1T (MoE)	1M	Huawei Ascend 950PR
GPT-5.4	Undisclosed	256K	Nvidia H200/B200
Claude Opus 4.6	Undisclosed	200K	Nvidia/Trainium
Gemini 3.1 Pro	Undisclosed	2M	Google TPUv5

Last verified: April 2026