The Tamesis
Chinese artificial intelligence startup DeepSeek on Friday unveiled its long-awaited new model, now adapted to run on Huawei’s Ascend chips.
The new V4 is divided into two versions, DeepSeek-V4-Pro and DeepSeek-V4-Flash, with the latter being “a more efficient and economical option” because of its smaller parameters.
DeepSeek says V4-Pro lags only behind the latest Gemini model in terms of the inference benchmark “world knowledge.”
Also on AF: Chinese young people conquer artificial intelligence and subsidize the establishment of one-person companies
The company said a “preview version” of the open source model is now available, but did not say when the final version would be released.
The model is also “optimized” for popular AI agent products such as Claude Code, OpenClaw, OpenCode and CodeBuddy, DeepSeek stated.
It also runs on chips made by Chinese tech giant Huawei, the company added.
Huawei, which has been subject to U.S. sanctions over national security concerns since 2019, said in a statement on Friday that its entire range of Ascend SuperPoD products supports DeepSeek’s V4 series.
Earlier this month, The Information reported report It is understood that Chinese technology giants including Alibaba Group, ByteDance and Tencent have placed large orders for Huawei’s upcoming chips in preparation for the release of V4.
The report quoted five people with direct knowledge of the procurement as saying that the total purchase volume was hundreds of thousands of units.
Having V4 run on Huawei chips contrasts with DeepSeek’s past reliance on U.S. chipmaker Nvidia Semiconductor and marks a breakthrough for China’s artificial intelligence industry.
“This is a big thing for China’s artificial intelligence industry,” said He Hui, director of semiconductor research at consulting firm Omdia. told Reuters.
“Huawei’s Ascend chip is China’s best domestic chip and can replace Nvidia, and its support for DeepSeek V4 shows that China’s top artificial intelligence models can now run on Chinese hardware.”
Most leading artificial intelligence models are trained and run on chips made by Nvidia. Huawei says its chips are used in part of the V4’s training process.
“Significantly reduced” costs
Hangzhou-based DeepSeek emerged in January last year with a generative artificial intelligence chatbot that shocked the world with its low-cost inference model (R1), whose capabilities are comparable to those of American competitors.
The announcement triggered a sell-off in AI-related stocks and a rethink on business strategy in what has been dubbed the industry’s “Sputnik moment.”
The chatbot’s performance is similar to ChatGPT and other top US products, but the company says its development required far less computing power.
DeepSeek’s artificial intelligence tools have been widely adopted by Chinese municipalities, medical institutions, as well as the financial sector and other enterprises.
This is partly due to DeepSeek’s decision to open source its system, exposing its inner workings, in contrast to the proprietary models sold by OpenAI and other Western rivals.
Experts say the arrival of V4 marks an “inflection point” in terms of hardware and cost.
The company said in a statement on social media platform WeChat that DeepSeek-V4 “has ultra-long context,” and praised it in a separate statement on X for being “world-leading… and greatly reducing computing (and) memory costs.”
V4 supports a context length of one million “tokens” (small components of text, including words or punctuation), making it comparable to Google’s Gemini.
Context length determines how much input the model can absorb to help it complete its task.
“This solves the long-standing problem of slower performance and higher costs associated with long context lengths and marks a real inflection point for the industry,” Zhang Yi, founder of technology research firm iiMedia, told AFP.
“For end users, this will bring broad, accessible benefits. For example, if very long context support becomes a standard feature, long text processing is expected to move beyond high-end research labs and into mainstream commercial applications,” he said.
V4-Pro has 1.6 trillion parameters and V4-Flash has 284 billion parameters, which refines the model’s decision-making capabilities.
US charges technology theft
Max Liu, a senior artificial intelligence industry analyst, told AFP that DeepSeek’s latest version was a “milestone” for the Chinese company.
“This is a good thing for the entire domestic AI industry. It can provide better models to domestic users and we can now look forward to more things – more products (and) a more competitive market,” he told AFP.
He added that if DeepSeek’s new model does match the performance of leading models from Western labs, “it would be as shocking as when DeepSeek first came out.”
Still, models like Deepseek have intensified the tech competition between China and the United States.
On Thursday, the White House accused Chinese entities of massive theft of artificial intelligence technology.
Michael Kratsios, Trump’s top tech adviser, posted on X: “The United States has evidence that foreign entities, primarily in China, are conducting industrial-scale distillation operations to steal American artificial intelligence.”
Distillation is a common practice in AI development, with companies often using it to create cheaper, smaller versions of their own models.
The US comments come ahead of next month’s summit between Donald Trump and Xi Jinping in Beijing.
China, meanwhile, called the White House’s claims “baseless.”
“The U.S. statement is completely baseless,” Chinese Foreign Ministry spokesman Guo Jiakun said at a press conference in Beijing. “This is a slander against the achievements of China’s artificial intelligence industry.”
- AFP, Additional editing by Vishakha Saxena, Information provided by Reuters
Also read:
Chinese AI circuit board company raises US$2 billion, becomes Hong Kong listed company of the year
China’s top streaming site criticized for AI actor ‘database’
Bytedance’s viral video model shows China’s growing control over AI
New DeepSeek model delayed online to test China’s AI chip strength
Nexperia’s Chinese subsidiary’s chips “nearly fully localized”
China beats U.S. again in global patent battle, Huawei maintains lead
China now requires chipmakers to use at least 50% domestic equipment
Nvidia CEO says U.S. export restrictions on AI chips are ‘flawed’ policy
China’s big technological secret: Dutch chip-making machine replicated in laboratory
Chinese AI firms form alliance with chipmakers to ditch foreign technology
China’s “tech economy” trumps its bleak growth prospects
Beijing and its billions in artificial intelligence
TSMC’s net profit jumps more than 58% as demand for artificial intelligence reaches new highs
Samsung Electronics expects profits to surge 755% due to AI boom
