DeepSeek introduced the new V4 model and crashed competitors' shares
Chinese startup DeepSeek has released a preview of its new large language model V4. It is open source and available for testing in the pro and flash versions.
Chinese startup DeepSeek has released a preview of its new large language model V4. It is open source and available for testing in the pro and flash versions.
As CNBC reports, the release of the V4 comes amid heightened domestic competition in China. Unlike the legendary R1 model, which shocked the world with its low cost and performance in early 2025, the current update is aimed at directly combating local giants like Alibaba and ByteDance.
The developers presented two main variations of the model:
-
DeepSeek-V4-Pro: has 1.6 TB of total parameters (49 billion active). According to the company, its performance is not inferior to the world’s leading closed-source models.
-
DeepSeek-V4-Flash: a lighter version with 284 billion parameters (13 billion active), focused on maximum speed and resource saving.
Analysts note that DeepSeek V4 offers significantly lower inference costs compared to competitors. «The preview of DeepSeek V4 is a major breakthrough. It demonstrates excellent agent capabilities at a significantly lower cost,» said Wei Song, principal AI analyst at Counterpoint Research.
The developers have optimized the model for integration with popular tools such as Claude Code from Anthropic, allowing V4 to be used to automate complex tasks in software development and knowledge analytics.
Although the market has become accustomed to DeepSeek’s aggressive pricing policy, the release of V4 has caused other AI players to fall in shares. On the Hong Kong Stock Exchange, MiniMax and Zhipu shares fell by 8%, and Manycore Tech fell by 9%. Investors fear that another dumping from the Hangzhou startup will once again force the market to reconsider the margins of the business.
Recall that DeepSeek became a real sensation in January 2025. Then the company proved that it was capable of training models at the level of industry leaders (OpenAI, Google) on outdated hardware and with a budget of up to $6 million. This refuted the thesis that leadership in AI necessarily requires tens of billions of dollars and the latest Nvidia chips, the supply of which to China and Russia is limited by US sanctions.
Read the country’s main IT news in our Telegram
Read the country’s main IT news in our Telegram
The mysterious Hunter Alpha AI model has intrigued developers: is this really a new blockbuster from DeepSeek?
The mysterious Hunter Alpha AI model has intrigued developers: is this really a new blockbuster from DeepSeek?
ByteDance’s AI video model — Seedance 2.0 — goes viral: Chinese already predict its success DeepSeek
ByteDance’s AI video model — Seedance 2.0 — goes viral: Chinese already predict its success DeepSeek
DeepSeek has found a new approach to training LLM models that could revolutionize the AI market again
DeepSeek has found a new approach to training LLM models that could revolutionize the AI market again
