Just now, DeepSeek's new model was suddenly launched, which is a Flash-tier model with Pro-level ambition.

Just now, DeepSeek has launched a limited-time internal beta test for its new model DeepSeek V4.1 Flash.

The official states that this is an intermediate version open for user testing, which adopts a new model architecture and natively supports multi-modality. According to the information provided by DeepSeek, the new model not only has stronger capabilities and faster generation speed, but also further reduces the usage cost. 

Users can now access the API without modifying the original base_url. As long as the model name is set to deepseek-v4.1-flash-expires-on-0910, it can be called directly. 

The billing standard during the test period is the same as that of DeepSeek V4 Flash, and each account supports a maximum of 20 concurrent requests. 

It is worth noting that V4.1 Flash is still a limited-time test version at present, and the "expires-on-0910" in the model name indicates that this version will automatically expire and go offline on September 10.

Judging from the name, it is a Flash model, but from DeepSeek's attitude, it obviously does not just want to make a faster and cheaper version. For example, in the feedback questionnaire opened along with the test, DeepSeek directly asked users: 

"Do you think this model can completely replace the DeepSeek V4 Pro online?"

The implication is that DeepSeek's expectation for V4.1 Flash is not to replace the previous generation of Flash, but to move closer to the capabilities of V4 Pro on the basis of the speed and price of Flash. 

As usual, we quickly accessed Claude Code, and then used last year's college entrance examination essay to simply test its Chinese long-form writing ability. 

In terms of actual performance, the opening, transition and conclusion of the article have clear designs, and the emotional progression is relatively natural. 

Compared with some model essays that are accustomed to piling up famous quotes and forcibly elevating themes, the narrative rhythm of V4.1 Flash is more human-like than previous versions. 

If the essay test examines whether it can express ideas well, then the following front-end cases examine whether it can turn ideas into real outputs. 

The second test requires the model to use only HTML, CSS and JavaScript to simulate an interactive instant camera. 

It only took 5 minutes for V4.1 Flash to deliver an answer with a mature understanding of the visual hierarchy of the web page. There is a clear spatial relationship between the camera shell, lens, shutter and photo outlet, and the light and shadow effects do not overshadow the core content. 

Next, we asked it to create a macOS-like web operating system in a single HTML file, which includes a text editor, terminal, Python environment, code editor, games, file manager, drawing tool and video editor. 

Faced with this classic case, the performance of V4.1 Flash can only be described as mediocre, without bringing too many surprises. 

But when we asked V4.1 Flash to create a sci-fi spaceship cockpit HUD, this case reflected its progress in UI and interaction details. 

It sets slightly different response speeds for different interface layers. The central crosshair is the most sensitive, the peripheral instruments are slightly slower, and the background moves with a smaller amplitude, thus creating a sense of depth between the cockpit glass, the attitude of the spaceship and the distant space. 

After several rounds of simple experience, the generation speed of V4.1 Flash is a major highlight, and the continuous generation is also relatively stable. For programming scenarios that require frequent trial and error, the actual value of this speed improvement is more easily perceived by users than the increase in benchmark scores. 

The current V4.1 Flash is still an intermediate test version, but if Flash can provide performance close to Pro while retaining faster response speed and lower call cost, the familiar playstyle of DeepSeek may be back. 

Bring stronger capabilities to more users at a lower price.

This article is from the WeChat official account "APPSO", author: Discover Tomorrow's Products, published by 36Kr with authorization.