A year after shaking up Silicon Valley with its technology, Chinese company DeepSeek has released preview versions of a new artificial intelligence model considered its flagship — and is already selling it as “the most powerful open source platform on the market”, in a direct message to rivals like OpenAI and Anthropic.
The startup presented the V4 Flash and V4 Pro series, saying that the models have cutting-edge performance in programming tests and bring important advances in reasoning and tasks performed by autonomous agents. According to the company, they arrive with architectural changes and optimization improvements, detailed in a post on Hugging Face. DeepSeek highlighted a technique called Hybrid Attention Architecture, which, according to the company, improves the model’s ability to remember questions and answers throughout long conversations. The context window has also been expanded to 1 million tokens — allowing you to send entire codebases or very large documents in a single prompt.
V4 comes more than a year after Hangzhou-based DeepSeek triggered a share sale that erased more than $1 trillion in market value with the launch of R1, an open model that mimics the human reasoning process. The R1 rivaled cutting-edge AI systems from companies like OpenAI, but it would have been developed at a fraction of the cost of American competitors.
Continues after advertising
On the Chinese stock market, shares of local chip manufacturers soared this Friday, with investors betting that the new model should sustain demand for semiconductors produced in China. In a post on WeChat, DeepSeek acknowledged that the service capacity of the V4 Pro line is still “extremely limited” due to the lack of computing infrastructure. The company’s expectation is that the cost of using the models will drop significantly after the entry into operation, in the second half of the year, of clusters equipped with Ascend 950 chips, from Huawei Technologies Co. The startup is also negotiating with Tencent and Alibaba its first round of funding from external investors.
The new family of models is another leap in scale and efficiency — two hallmarks of DeepSeek’s trajectory — and increases competitive pressure on rivals. After R1, technology giants and investors came to rethink the logic of pouring billions of dollars into AI projects. This movement, however, has lost momentum: American technology companies are expected to invest around US$650 billion in 2026 in AI infrastructure and data centers alone.
DeepSeek’s 1 trillion parameter system uses the Mixture-of-Experts technique, which activates only a portion of the model’s “experts” in each task, activating a maximum of 37 billion parameters per request. Therefore, the inference cost is well below that of other frontier models with a similar scale.
According to the company, the DeepSeek-V4 Pro costs just a fraction of what the main US laboratories charge. Input tokens — the user-submitted text — come in at $1.74 per million, while output tokens — the model’s response — cost $3.48 per million. On Claude Sonnet 4, Anthropic’s top-of-the-line model, a million input tokens cost US$3, and an output token costs US$15.
The combination of architecture and techniques puts DeepSeek in direct confrontation with the latest models from OpenAI, Google and Anthropic. This Friday, the startup stated that it outperforms systems such as OpenAI’s GPT-5.2 in standardized benchmarks, but admitted that the V4 is still 3 to 6 months behind the most advanced models on the market.
Even so, the company makes a point of saying that it is not just playing the game of “brute force”, but rather one of cost reduction. V4 was designed to run on cheaper infrastructure, which could further tighten the grip on other Chinese AI companies, such as MiniMax.
Continues after advertising
“Minimax and Zhipu, as independent model providers, will always be exposed to competition, especially from large internet or cloud platforms, which have much greater reach and distribution,” said Vey-Sern Ling, managing director of Union Bancaire Privée. “Ultimately, the difference in performance between models tends to be imperceptible to most users.”
DeepSeek’s breakthrough last year sparked an AI rush in China, with giants like Alibaba and Baidu pouring cheap artificial intelligence services onto the market. Rivals like ByteDance, Zhipu and Minimax rushed to review and relaunch their models in the weeks leading up to April in an attempt not to fall behind.
With stardom also came scrutiny. Leaders of technology companies and US authorities accuse DeepSeek of using prohibited techniques and hardware to train its systems.
Continues after advertising
OpenAI and Anthropic have already stated that DeepSeek resorted to “distillation” — a process in which the output of other AI models is used to train a system with similar capabilities. The two American companies say they detected this type of use coming from the Chinese startup, and OpenAI began raising the issue behind the scenes shortly after the launch of R1.
There is also a suspicion that DeepSeek has access to Nvidia AI chips that are on the list of components vetoed for China. According to people familiar with the matter, the American government believes that the startup used Nvidia’s Blackwell processors — effectively banned for the Chinese market — in a data center in the Inner Mongolia region.
© 2026 Bloomberg L.P.
Continues after advertising
