Summary of the Key News Points in Plain Language
Domestic AI large-model manufacturer DeepSeek recently launched a very unusual “limited-time 2-day beta test” campaign: On September 8th, they released the new V4.1 Flash model for a small group of developers to test, and on September 10th, the model was automatically taken down. The test results showed that the new model was 5-6 times faster than the previous version. However, developers did not experience the claimed “lower costs”; instead, they felt as if their money was being deducted instantly due to the rapid output of results from the AI.
This seemingly playful short-term test is actually part of DeepSeek’s rapid iteration pace over the past 40 days, during which they have released five new products. Their large-model capabilities have dropped to 15th place globally, surpassed by domestic competitors such as Zhipu, Kimi, and Qianwen. Now, DeepSeek is trying to regain their position in the top tier of the global AI industry by combining a new model architecture with a comprehensive set of developer tools.
---
Detailed Analysis of Each Point
1. This “beta test” is not a real test; it’s a deliberate pre-release to gain an advantage
Many people might think DeepSeek is being too stingy for only offering a 2-day beta with a concurrent user limit of 20, but this is not a proper beta test at all. Normally, companies would conduct beta tests with a large number of users over several months to identify and fix bugs and prepare for the official release. Offering a limited period and a concurrent user limit that makes it impossible to conduct actual business operations is not typical of a beta test. Essentially, DeepSeek is using a form of “hunger marketing” among developers: developers don’t even need to modify the interface; they can simply use the new model by changing its name. The 2-day limit creates a sense of urgency, encouraging them to test the new model and share their experiences on social media. This strategy spreads the message that the new model is much faster than the old one without incurring any advertising costs, and the impact is ten times more effective than issuing a formal press release.
2. The claimed 6-fold speed increase is not a minor upgrade; it represents a shift in the competition within the AI industry
In the past, the competition focused on which model could solve problems more efficiently or write code more accurately, which meant comparing who was smarter. Ordinary users could hardly notice a 1% or 2% difference in performance. This time, DeepSeek is focusing on speed: generating SVG images is 6 times faster, and searching through long documents for information is 5 times faster. For example, what used to take half a minute (like waiting for an AI to generate a poster) now happens instantly. This significant improvement is noticeable to all users.
The company has also made it clear that this is not just a minor adjustment of parameters but a complete overhaul of the model architecture, similar to a smartphone upgrading from 4G to 5G. The entire AI industry will soon follow this trend of focusing on speed, and users will experience less lag and waiting times when using AI services.
3. Why do developers feel the costs are higher even though the company claims they are lower?
The difference in perception here is easy to understand: You can think of the AI’s billing unit (tokens) as “the amount of text processed by the AI.” When the company says the costs are lower, it doesn’t mean the actual cost to developers has decreased. For example, imagine taking a taxi: if the AI was slow, it would cost the same amount to travel 10 kilometers as it does now, even though the speed has increased significantly. The company’s claim of lower costs actually refers to their own reduced expenses. For instance, if a server could process 1000 words per second and the cost was 5 yuan, with the new model, it can process 6000 words per second, reducing the cost per 1000 words to just a few cents. The company is making more profits, but they haven’t yet passed this benefit on to developers.
4. Releasing five products in 40 days: DeepSeek is focusing on building an ecosystem, not just adding features
Many wonder why DeepSeek is updating so frequently. They are not just fixing individual model shortcomings but are building a strong developer ecosystem. Previously, they only provided the “large-model API,” which was like selling developers just an engine. To create an AI application that could recognize images or search for information, developers had to find and assemble other components themselves, which was a high barrier. Now, DeepSeek provides everything needed, from the underlying model to multi-modal capabilities and a full set of development tools, essentially giving developers a ready-made solution. Developers are more likely to stick with DeepSeek because of this convenience, making it easier for them to launch their own AI products. Once they get used to this ecosystem, they are less likely to switch to other platforms, thus keeping users firmly tied to DeepSeek’s platform.
5. The 2-day test reflects a complete shift in the domestic AI landscape
The AI industry has moved beyond the initial phase where the first company to release a large model could reap significant benefits. The current competition focuses on two key aspects:
- Hidden cost-effectiveness: DeepSeek’s lower underlying processing costs mean they can lower prices to attract customers, leaving competitors at a disadvantage if their costs are higher.
- Developer retention: 90% of new AI applications are created by small developers. The company with the most comprehensive and user-friendly toolchain will attract the most developers, which is essential for building a large ecosystem. DeepSeek is using speed and frequent updates to gain a competitive advantage, aiming to bypass the traditional focus on performance metrics. In the next six months, the domestic AI industry will enter a new phase where models will be faster and more affordable, and users will experience a significant improvement in their AI experiences.