Chinese artificial intelligence startup DeepSeek has released a preview of its long-awaited V4 model, marking its most significant update since it upended the global AI industry more than a year ago, and signalling a deliberate push toward hardware independence from American chipmakers.
The release comes over a year after DeepSeek’s R1 reasoning model stunned the global AI industry in January 2025, turning the Hangzhou-based startup from a little-known research team into China’s best-known AI company almost overnight and raising fundamental questions about how much computing power frontier AI actually requires.
DeepSeek has released two open-weight models in preview: DeepSeek V4, a mixture-of-experts model with 284 billion parameters, and DeepSeek V4-Pro, a far larger model with 1.6 trillion parameters trained on 33 trillion tokens. Both are available through the DeepSeek application programming interface (API), its web service, and the Hugging Face platform.
The V4-Pro version claims to outperform all open-source models in world-knowledge benchmarks, trailing only Google’s Gemini-Pro-3.1 and falling slightly behind OpenAI’s GPT-5.4. It excels in coding, reasoning, and science, technology, engineering and mathematics (STEM) tasks, with a one-million-token context window matching top-tier closed models. These performance claims are self-reported and independent verification is ongoing.
On pricing, DeepSeek V4 is offered at 14 US cents per million input tokens, while V4-Pro is priced at $1.74 per million input tokens, compared to OpenAI’s GPT-5.5 at $5 per million input tokens.
The hardware dimension of the release drew particular attention. V4 is DeepSeek’s first model optimised for domestic Chinese chips, including Huawei’s Ascend processors, turning the launch into a test of whether China’s AI industry can loosen its dependence on US chip giant Nvidia. Huawei confirmed its Ascend supernode products would offer full support for DeepSeek V4.
However, the shift is not yet complete. DeepSeek appears to have adapted only part of V4’s training process for Chinese chips, with the model using domestic hardware for inference while potentially still relying on Nvidia chips for portions of training. The technical report does not confirm whether all key features were adapted to domestic hardware.
Shares in Semiconductor Manufacturing International Corporation (SMIC), the Chinese chipmaker that manufactures Huawei’s Ascend AI processors, jumped 10 percent in Hong Kong trading following the announcement.
The release landed amid heightened geopolitical tension. The launch came a day after the White House accused China of industrial-scale theft of US AI intellectual property. OpenAI and Anthropic have also accused Chinese AI developers, including DeepSeek, of conducting illicit distillation attacks, while China’s foreign ministry dismissed the claims as groundless.
