DeepSeek V4 has achieved the performance of billion-dollar closed-source systems as a free open-source (source code publicly available for anyone to use and modify) model. This signals that the moat in the large model domain is shifting from "compute scale" to "engineering efficiency."
What this is
Data disclosed by Two Minute Papers this week shows that DeepSeek's new V4 model matches top-tier closed-source models trained on billions of dollars of compute across multiple core benchmarks—and it is entirely free. This wasn't achieved by blindly stacking GPUs, but through smarter algorithmic architectures and training strategies, squeezing maximum efficiency out of every ounce of compute. We note that while the industry is still debating whether the Scaling Law (the law stating model performance improves with increased compute) is slowing down, DeepSeek has provided a more pragmatic path: achieving the same results with less money.
Industry view
The market generally views this as another major victory for the open-source camp. It means startups no longer need to be locked into expensive closed-source APIs (Application Programming Interfaces: channels for software interaction), and small-to-medium teams can invoke top-tier capabilities at extremely low costs. However, we are concerned about the underlying risks: on one hand, the "free" model creates massive cash flow pressure for the provider; training compute costs haven't disappeared, they've simply been shifted or subsidized. On the other hand, when foundational model capabilities are easily leveled, homogeneous competition at the application layer becomes inevitable, potentially devolving into an industry-wide price war that compresses the space for long-term innovation.
Impact on regular people
For enterprise IT: The barrier to deploying top-tier AI has dropped significantly. Localized deployment (installing models on the company's own servers) is no longer a high-cost game, making data privacy and security easier to control.
For the individual workplace: The barrier to accessing powerful AI capabilities has been flattened. Those who can combine free tools to create solutions will hold a greater cost advantage than those relying solely on expensive subscriptions.
For the consumer market: The trial-and-error costs for application developers are reduced. In the future, we expect to see the emergence of more low-priced or even free intelligent software services.