Crypto

DeepSeek V4 Launches on Huawei Chips


Chinese AI startup DeepSeek released a preview of its V4 model on April 24, explicitly optimized for Huawei’s Ascend chip platform, on the same day the White House accused China of running industrial-scale campaigns to copy and steal American frontier AI systems.

Summary

  • DeepSeek released V4-Pro and V4-Flash preview models on April 24, both built to run on Huawei’s Ascend AI accelerators, marking a strategic pivot away from Nvidia hardware.
  • V4-Pro carries 1.6 trillion total parameters and trails only Google’s closed-source Gemini-Pro-3.1 on world knowledge benchmarks among open-source models.
  • The launch arrived hours after a White House OSTP memo accused Chinese entities of using tens of thousands of proxy accounts and jailbreaking techniques to distil proprietary US AI models.

DeepSeek released preview versions of its V4-Pro and V4-Flash models on April 24, both explicitly designed to run on Huawei’s Ascend AI chip platform. Huawei confirmed that its full Ascend supernode product line now supports the DeepSeek V4 series. The launch came one day after a White House memo accused China of coordinated, industrial-scale intellectual property theft from American AI laboratories, with DeepSeek cited repeatedly in prior US accusations as a company that distilled proprietary OpenAI and Anthropic models.

DeepSeek V4 Huawei Chips Launch Inverts the Nvidia Dependency Narrative

DeepSeek’s earlier V3 model was trained on Nvidia hardware, which drew accusations from Washington that the company breached US export controls to acquire advanced Nvidia chips. The V4 announcement inverts that positioning entirely. Huawei is front and center as both collaborator and primary deployment target, while Nvidia is absent from the technical documentation. V4-Pro operates on a mixture-of-experts architecture with 1.6 trillion total parameters and 49 billion active parameters. V4-Flash uses a smaller 284 billion total parameter configuration with 13 billion active parameters designed for cost efficiency. DeepSeek said V4-Pro outperforms every other open-source model on world knowledge benchmarks and trails only Google’s closed-source Gemini-Pro-3.1. The launch also includes a lower-cost flash variant, the same two-tier pricing pattern DeepSeek used to undercut Western labs with V3. As crypto.news reported, Nvidia CEO Jensen Huang had warned the week prior that China possesses the infrastructure and hardware necessary to rival the capabilities of US frontier AI models.

The White House Memo That Landed the Day Before

Michael Kratsios, director of the White House Office of Science and Technology Policy, circulated a memo on April 23 accusing Chinese entities of running “industrial-scale campaigns to distil US frontier AI systems” using tens of thousands of proxy accounts to evade detection and jailbreaking techniques to expose proprietary model information. Distillation is the process of training smaller AI models on the outputs of larger ones, a method that allows a lab to approximate frontier capability without paying the full training cost. The memo stated the administration will share information with US AI companies about distillation campaigns and explore accountability measures against foreign actors. As crypto.news documented, the US AI industry has been navigating compounding geopolitical pressures in 2026, with Chainlink and other AI-adjacent crypto assets already showing sensitivity to US-China tech tensions earlier in April.

What DeepSeek V4 on Huawei Means for the US Chip Export Strategy

The V4 Huawei pivot is a direct strategic response to US export controls. By demonstrating that a model approaching frontier performance can now be trained and served on Huawei Ascend silicon, DeepSeek and China are challenging the assumption that semiconductor restrictions can meaningfully slow Chinese AI development. Commerce Secretary Howard Lutnick confirmed separately on April 23 that no Nvidia advanced AI chip shipments had actually gone through to China despite the January conditional approval, a disclosure that deepens the question of what hardware DeepSeek actually used for V4 training. DeepSeek has not confirmed this, and detailed technical specifications for V4 have not been independently verified. As crypto.news tracked, Nvidia disclosed $5.5 billion in expected charges in April 2025 when the US government required export licenses for H20 chips sold to China, a policy backdrop that made Huawei’s role in V4 commercially inevitable for DeepSeek’s long-term supply chain.

The Chinese Embassy in Washington called the White House distillation accusations “baseless,” and Beijing’s foreign ministry urged the US to “abandon biases” and said more scientific exchange, not less, was needed.


Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button