DeepSeek V4: First Chinese Frontier Model Built for Huawei Chips

China's DeepSeek released a preview of its V4 model on Friday, and the headline is not the benchmarks. It is what the model runs on. V4 was co-developed with Huawei from the beginning. Parts of the training ran on Huawei's Ascend chips. The full inference stack now runs on Chinese hardware, end to end.

For the first time, a top-tier Chinese AI model is fully decoupled from Nvidia silicon.

The model

V4-Pro outperforms every other open-source model on world-knowledge benchmarks and trails only Google's closed-source Gemini-Pro-3.1, DeepSeek said. SCMP described the architecture as "world-leading" on efficiency. Bloomberg, citing CCTV, reported that the V4 release was held back by months specifically to complete the migration onto Huawei chips, not for capability tuning.

He Hui, semiconductor research director at Omdia, summed up the read in plain terms.

"This is a big deal for China's AI industry. Huawei's Ascend chips are the country's best homegrown alternative to Nvidia, and supporting DeepSeek V4 shows that top Chinese AI models can now run on Chinese hardware."

The technical bet

Most leading AI models, including OpenAI's, Anthropic's and DeepSeek's own previous releases, were trained primarily on Nvidia GPUs. V4 changes that contract. Huawei contributed compute on its Ascend platform during training. The deployment stack is fully native to Ascend. Chinese cloud providers, the central buyers of these models, can now run V4 on domestic chips without any Nvidia in the loop.

That matters because the latest US House export-control bills, advanced on April 22, would tighten chip flows to China further. If the bills pass, Chinese AI infrastructure that depends on Nvidia hardware becomes an open vulnerability. Models that run native on Ascend are insulated.

The market read

Tencent and Alibaba are in talks to invest in DeepSeek at over $20 billion this week. Chinese chip stocks rallied on the V4 news, with investors rotating into Cambricon, Hygon and Huawei suppliers. Bloomberg's frame from a Friday note was harsher: "DeepSeek's long-awaited new model fails to narrow US lead in AI." The benchmarks back that view on raw capability. Gemini-Pro-3.1 still leads on world knowledge, and frontier reasoning still belongs to closed US labs.

But the framing misses the strategic point. V4 was never about closing the benchmark gap to Google. It was about closing the hardware gap to Nvidia. On that front, the work is shipping.

Sources: SCMP, DealStreet Asia, Nikkei, Bloomberg