DeepSeek launches V4 model adapted for Huawei AI chips

• DeepSeek launches V4 for Huawei AI chips.
• Targets long-context agent tasks.

DeepSeek launched a preview version of its V4 artificial intelligence model on Friday, with the Chinese startup adapting the system for Huawei chip technology, according to Reuters.

The release is a closer technical link between DeepSeek and Huawei after DeepSeek’s earlier V3 and R1 models were trained on Nvidia chips. Huawei said its chips were used for part of V4-Flash’s training and that V4 is supported on its Ascend 950-based supernode clusters.

Training generally requires more compute than inference, extending Huawei’s role beyond deployment support. Inference refers to the stage where a completed model generates responses.

In maximum reasoning mode, DeepSeek said V4 Pro outperformed all open-source models, though it still trailed frontier closed-source systems like Google’s Gemini 3.1 Pro and OpenAI’s GPT-5.4 in some areas, according to a DeepSeek paper released with the model.

The performance claims come from DeepSeek’s paper and are yet to be tested by independent researchers or developers. Daniel Dewhurst, an AI engineer who tested V4 after its release, told Reuters the preview looked the results should be treated cautiously until independent evaluations and wider developer testing are available.

An important change from earlier DeepSeek releases is that V4 was adapted for Huawei’s most advanced Ascend AI chips. Huawei’s reported role in part of V4-Flash’s training places its chips in part of the model development process as well as deployment.

Nvidia chips are used to train and deploy most closed-source AI models. DeepSeek’s use of Huawei technology comes as US export controls limit China’s access to some advanced AI processors, while Beijing continues to support domestic chip development.

Nvidia CEO Jensen Huang has warned that US restrictions could weaken the company’s developer base in China. Speaking on a podcast this month, Huang said DeepSeek launching first on Huawei would be “a horrible outcome” for the US.

The V4 model quickly became the top trending model on Hugging Face. Lewis Tunstall, a machine learning engineer at Hugging Face, said it was the fastest model to reach that position. He said V4 performs well on long and complex text tasks and operates at a lower cost than competing top models. He also noted that the model does not support multiple modalities, like images and video.

V4 targets agent workloads

DeepSeek said V4 is designed to work with agent frameworks, including Claude Code and OpenClaw. Such workloads require models to retain instructions, tool outputs, and task history in longer sequences, making context length and compute cost more relevant.

Both V4 versions support a one-million-token context window, which allows it to process large documents, codebases, or extended task histories in a single prompt, a feature relevant to coding and agent workflows.

V4 comes in two versions: the higher-end Pro model and the cheaper, lighter Flash version. Flash delivers similar reasoning ability in some areas but runs faster and at a lower cost than Pro. It has weaker world knowledge and lower performance on more demanding agent-based tasks.

DeepSeek said Pro can cost up to 12 times more than Flash because of constraints in compute capacity. The company said Pro pricing could fall once Huawei Ascend 950 supernodes are deployed at scale in the second half of the year.

Export controls frame the launch

The launch comes as DeepSeek faces scrutiny in the US over its technology sources and data practices. The company has acknowledged using Nvidia chips but has not said whether those chips were subject to export restrictions. It has also said it has not intentionally used synthetic data generated by OpenAI.

A day earlier, the White House accused China of stealing intellectual property from US AI labs. The V4 launch also comes ahead of US President Donald Trump’s planned visit to Beijing next month to meet Chinese President Xi Jinping.

The Trump administration approved the sale of Nvidia’s H200 chips to China in January. Sources cited by Reuters said shipments have been held up by disagreements between China and the US over sales terms.

Chinese chip stocks rose after the DeepSeek release, with Huahong Semiconductor gaining 15% and SMIC rising 10%. Nvidia shares also rose, supported by Intel’s stronger revenue and profit forecast.

DeepSeek’s models are used on many international platforms, despite restrictions by several Western and Asian governments, with various governments citing data privacy concerns, and banning some institutions and officials from using DeepSeek.

In China, DeepSeek faces stronger competition from domestic AI companies after gaining national attention last year. Shares of rivals Zhipu AI and MiniMax fell 9% after the V4 release.

Lian Jye Su, chief analyst at Omdia, told Reuters that the partnership shows DeepSeek models can deliver similar performance on Huawei and Nvidia hardware. He also said Huawei still trails Nvidia technologically, and that moving developers away from Nvidia’s ecosystem remains difficult.

DeepSeek is owned by China’s High-Flyer Capital Management. Reuters reported, citing The Informationthat Alibaba and Tencent were in talks to invest in DeepSeek at a valuation above US$20 billion, though Reuters said it could not independently verify the reports.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.