Portuguese 
English 
Spanish 
4 min of reading 
0 comments 
DeepSeek V4 expands the global artificial intelligence dispute by combining a 1 million token context window, gains in coding, reasoning, and agent-oriented tasks, in addition to a strategy focused on lower costs, national hardware, and open-source models to pressure competitors from the United States.
DeepSeek unveiled preview versions of its V4 AI model and entered a new stage of the global artificial intelligence dispute, directly targeting American platforms at a time of accelerating industry growth. The Chinese company is betting on a 1 million token context window, gains in coding, reasoning, and agent-oriented tasks, in addition to a strategy focused on reducing operational costs.
The launch occurs in a week marked by the advancement of rivals from the United States and accusations from the White House against China regarding large-scale copying of American AI systems. In this scenario, DeepSeek seeks to expand its space with the V4 Flash and V4 Pro series, which arrive with architectural updates, optimization improvements, and a declared focus on efficiency.
DeepSeek bets on a 1 million token context window
One of the main elements of the V4 model is the so-called Hybrid Attention Architecture, presented as a way to improve context retention in long conversations. The technology also seeks to reduce memory loss in prolonged interactions, a relevant point for more complex uses of artificial intelligence.
-
While you sleep, an artificial intelligence trained with 600,000 hours of sleep can predict 130 diseases — including Parkinson’s and cancer — years before the first symptoms.
-
New artificial intelligence technology uncovers hidden ocean currents and drastically changes predictions about the planet’s climate.
-
Netflix announces TikTok-style vertical feed and expands the use of AI in recommendations, content creation, and advertising, revealing a new strategy to transform the streaming experience.
-
As technology giants close increasingly powerful and strategic artificial intelligence models, Brazil risks being left out of the era of superintelligent AI and losing competitiveness in critical sectors such as the economy, defense, and innovation.
DeepSeek has also incorporated support for a 1 million token context window, which allows entire codebases or extensive documents to be inserted into a single prompt. This capability can change software development and business analysis routines, especially in tasks that depend on large volumes of information.
The V4 Flash and V4 Pro series were presented with solid performance in comparative tests involving systems from Anthropic, Google, and OpenAI. The company itself recognized, however, that version 4 is still three to six months behind the most advanced models, although it highlights flexibility in cost and implementation.
Efficiency becomes a central axis of the strategy
DeepSeek maintains efficiency as one of its main differentials in the dispute with rivals from the United States. The trillion-parameter system uses a Mixture of Experts approach, which activates only a fraction of the parameters for each task.
This operation reduces inference costs compared to traditional models, where all parameters are usually activated with each request. The technical choice reinforces the company’s attempt to compete in performance without relying on such an expensive operational structure.
The pressure for efficiency gains weight because AI systems are becoming increasingly expensive to operate. DeepSeek tries to position itself as an alternative in a market where performance, cost, and scale have become decisive factors for companies and developers.
Chinese chips enter the center of the dispute
DeepSeek’s models were also designed to run on national hardware, which increases the relevance of Chinese infrastructure in the company’s advancement. The expectation is for a further drop in costs when clusters equipped with Huawei Technologies Co.’s Ascend 950 chips come into operation later this year.
This change could reduce dependence on US chip manufacturers and strengthen China’s AI infrastructure. The strategy emerges amidst the dispute for computational capacity, considered essential for training, operating, and scaling more advanced models.
Market reaction was swift after the announcement. Shares of Semiconductor Manufacturing International Corp. and Hua Hong Semiconductor rose, while shares of competing AI companies declined.
The movement indicates an investor bet on the growth of demand for chips manufactured in China. DeepSeek, however, stated that the service capacity of the V4 Pro series remains limited by computational resource constraints.
Negotiations indicate expansion plans
The company is also in talks with Tencent Holdings Ltd. and Alibaba Group Holding Ltd. for its first round of funding. The negotiations signal infrastructure expansion plans, at a time when computational capacity limits part of the V4 Pro’s offering.
The launch of the V4 version follows the R1 model, which stirred the artificial intelligence market and prompted a re-evaluation of investments in cutting-edge systems. DeepSeek stated that R1 delivered competitive performance at a fraction of the cost of leading American models.
Since then, the debate over investments has regained momentum. Technology companies in the United States are expected to invest approximately US$650 billion in 2026 in AI infrastructure and data centers, attempting to balance performance gains and long-term costs.
DeepSeek increases pressure on closed models
DeepSeek states that version 4 deepens the strategy initiated with R1, with advancements in scalability and efficiency. The company also continues to position open-source models as alternatives to closed systems, attracting developers and companies seeking more control over their tools.
The launch, however, occurs under scrutiny. American authorities accused DeepSeek of using restricted chips, while Anthropic alleged misuse of its Claude system.
DeepSeek did not disclose training costs or hardware details for version 4. Nevertheless, V4 reinforces the dispute around lower costs, scalable performance, and hardware flexibility, points that place the Chinese company at the center of the next phase of global competition in artificial intelligence.
With information from Interesting Engineering
Be the first to react!