At the beginning of 2025, Deepseek surprisingly penetrated the AI world and caused a stock market earthquake with its appearance. Now the Chinese AI startup is following up with the long-awaited model Deepseek V4 in the Pro and Flash variants – but the big shock is unlikely to happen this time.
Just like the previous models, V4 is also open source. With the generation change, Deepseek relies on a completely new architecture and now works with 1.6 trillion parameters – more than twice as many as the previous model V3.2.
Deepseek V4: Improvement especially in the area of coding
In a research report on the new model published on April 24, Deepseek highlights “significant improvements in the thinking and action capabilities of the new model V4.” This is reported by CNN, among others. This is intended to enable autonomous action, such as writing code.
Editorial recommendations
${content}
${custom_anzeige-badge}
${custom_tr-badge}
${section}
${title}
In the version preview, Deepseek focuses on “top performances” in coding benchmarks and improved reasoning in three levels. The startup also states that V4 outperforms “all current open source models in math, STEM and programming.” When it comes to general knowledge, only Gemini 3.1 Pro is ahead.
The new model can also process up to one million tokens of context, while at the same time the computing and storage effort has been “drastically” reduced.
The colleagues at Heise state that experts place the actual “performance around three to six months behind the absolute top models on the market and not at eye level.”
Version update at Deepseek: The big shock is unlikely to happen
When Deepseek first came onto the market with its R1 model at the beginning of 2025, the competition was unprepared – and the prices of US AI stocks fell noticeably as a result. Now that Deepseek is no longer unknown in the AI race, the reaction is likely to be more cautious, experts estimate.
“The R1 shocked US markets because no one expected a Chinese model to be competitive at this level. The V4 is just a continuation of the same trend, and trends don’t generate headlines as much as shocks,” Mornigstar senior equity analyst Ivan Su told CNN.
Meanwhile, Deepseek remains unbeatable when it comes to price. In the most expensive variant (Pro), Deepseek V4 costs, according to benchmark information, only half of what OpenAI estimates for GPT 5.5 for comparable coding tasks.
Partnership with Huawei and focus on domestic chips
In order to cover the computing needs of V4 despite the scarce and expensive Nvidia and AMD chips, Deepseek worked with Huawei, among others, on the new model. According to its own statement, the Chinese technology group is supporting the AI startup with its “Supernode” technology, which combines large clusters of Ascend 950 chips.
The focus on Chinese chips from Huawei and Cambricon could ultimately work to Deepseek’s advantage, as an analyst at market research firm Counterpoint Research explains: “This will allow AI systems to be developed and deployed without relying solely on Nvidia. That’s why V4 could ultimately have an even greater impact than R1 – accelerating domestic adoption and contributing to faster global AI development overall.”
According to Heise, additional large Huawei Ascend 950 clusters will be added in the second half of 2026 to compensate for capacity bottlenecks, which Bloomberg (paywall) reported directly at the start of the V4 model.
Top Article
${content}
${custom_anzeige-badge}
${custom_tr-badge}
${section}
${title}
