China’s DeepSeek has introduced its next-generation artificial intelligence model, V4, while Huawei quickly pledged full support through its domestic chip ecosystem.
The announcement highlights not just a new model but also a broader push to reduce reliance on foreign technology.
DeepSeek released two versions of its new system. The larger V4-pro model has 1.6 trillion parameters, making it the company’s most powerful model to date. The smaller V4-flash model has 284 billion parameters.
In simple terms, more parameters usually mean stronger reasoning and better performance, but they also require far more computing power.
READ ALSO: MIT Quantum Sensor Reads Multiple Signals Together: Changing How We Measure Reality
The company says both models can handle up to 1 million tokens at once. This is a major jump from its previous limit of 128,000 tokens. A larger context window allows the AI to process longer documents, more detailed prompts, and complex tasks in a single go. DeepSeek claims it achieved this at world-leading cost efficiency.
The company also said its V4 models can compete with leading closed-source systems from OpenAI and Google DeepMind. Unlike those systems, V4 is open source, allowing developers to study and build on its architecture.
Shortly after the launch, Huawei confirmed that its Ascend chips and supernode systems will fully support the V4 models. The company shared more details during a livestream, showing how its hardware can handle model inference tasks.
Another Chinese chipmaker, Cambricon Technologies, also announced compatibility with the new models. Analysts see this as an important signal. They believe DeepSeek’s focus on domestic chip support could speed up the adoption of Chinese hardware across the AI sector.
WATCH ALSO: Humanoid robot folds towel autonomously after observing human
“The release of V4 clearly highlights compatibility with domestic chips,” analysts at Huatai Securities said in a note. “We expect rapid improvements in local graphics processing capabilities and wider adoption this year.”
Despite its power, the V4-pro model is too large to run on regular consumer devices. It requires advanced infrastructure, which remains limited. DeepSeek admitted that current performance is constrained by the availability of computing resources.
However, the company expects this to improve later this year. It said prices will drop significantly once Huawei’s Ascend 950PR systems begin shipping at scale.
The smaller V4-flash model, meanwhile, is positioned as one of the most affordable advanced AI models available. Its pricing matches DeepSeek’s earlier V2 model, making it attractive to developers seeking high performance at a lower cost.
DeepSeek did not reveal the hardware used to train V4. This is a shift from its earlier transparency, when it disclosed that it had used Nvidia H800 chips in previous models. The silence comes amid rising scrutiny from the US.
Before the launch, US officials accused DeepSeek of using restricted Nvidia chips, including advanced Blackwell processors. The company has not confirmed or denied these claims.
READ ALSO: Cornell Student Develops Ultra-Precise Microchips to Boost Tech Reliability
Still, DeepSeek’s technical report mentions custom kernels designed to work with both Nvidia and Huawei chips. This suggests a flexible approach that can adapt across different hardware platforms.
The ripple effects were immediate in financial markets. Shares of Semiconductor Manufacturing International Corporation rose more than 8 percent in Hong Kong trading. Hua Hong Semiconductor gained over 12 percent, while Hygon Information Technology saw nearly a 10 percent increase.
Analysts believe the new model will drive demand across China’s entire chip supply chain from GPUs to CPUs.
Despite speculation, V4 remains a text-only model. Many expected multimodal capabilities, such as the ability to process images and video. DeepSeek said it is still working on adding those features.
The company’s previous release, R1, had already drawn global attention. It narrowed the gap between Chinese and US AI systems to just months. Its reported training cost of $6 million also shocked the market, contributing to a massive one-day drop in Nvidia’s valuation.
WATCH ALSO: China’s new massive battle tank takes part in live-fire test
This time, DeepSeek is more cautious. It has not shared details about training costs or infrastructure. Still, the company is moving steadily forward.
“We follow long-term thinking,” DeepSeek said. “We move step by step, learning through testing and reflection, and aim to get closer to artificial general intelligence.”
With strong backing from Huawei and growing support from domestic chipmakers, V4 is more than just another AI model. It reflects a deeper shift in how and where the future of AI could be built.
