DeepSeek adds image and video recognition to its main chatbot

A Chinese artificial intelligence company has added image and video recognition to its main chatbot.

At the same time, local chip makers showed they can now match the fast launch support that used to be an American strength.

DeepSeek, a company based in Hangzhou, quietly added a new feature called “image recognition mode” to its chat platform. This new mode joins two other modes the company launched earlier this month: “expert” and “flash.”

The new feature allows the chatbot to understand photos and videos, not just text.

This brings it in line with other major AI chatbots that have offered similar abilities for some time.

Chen Xiaokang, who heads DeepSeek’s multimodal team, said the tool was first tested with a small group of users on both the website and the mobile app.

Chen Deli, a senior researcher at the company, celebrated the launch with a short post that referred to the company’s logo: “The little whale can now see.”

The image and video feature came out just a few days after DeepSeek released a preview of its newest flagship model, DeepSeek-V4, and made the model weights available for anyone to download and use.

V4 is not one model but two. The first one, DeepSeek-V4-Pro, has 1.6 trillion parameters and is designed for difficult tasks that need complex reasoning and multi-step automated workflows.

The second one, DeepSeek-V4-Flash, is built to handle a large number of requests at a lower cost. Both models support a context window of one million tokens.

They also use a hybrid attention design that the company says reduces computing power and memory needed during inference.

Chinese chip makers hit a new milestone

What caught the attention of many industry observers was not just the model itself, but what happened on the day it launched.

Four Chinese chip companies, Huawei Ascend, Cambricon, Hygon Information, and Moore Threads, all confirmed that their hardware worked perfectly with V4 from the very first day it was released.

This kind of same-day support, where a new model runs smoothly on non-NVIDIA chips right at launch instead of weeks or months later, had previously been almost impossible outside of Nvidia’s own ecosystem.

Huawei’s Ascend chips, including the A2, A3, and 950, support both V4-Pro and V4-Flash.

The company said its Ascend 950 chip uses fused computing processes and parallel processing streams to make inference faster.

Cambricon finished its adaptation using the open-source vLLM inference framework and shared its code on GitHub.

Hygon said it carried out deep model optimization on its DCU platform to create a smooth path from model release to actual use.

Moore Threads worked with the Beijing Academy of Artificial Intelligence to run V4 on its MTT S5000 card using the FlagOS software stack.

Industry watchers say this coordinated launch represents a real change.

For years, chips made outside Nvidia’s ecosystem would take months to support a major new model. Getting eight different domestic chipsets to work on day one is a significant milestone.

The bigger picture: cost and independence for Deepseek

Observers believe the bigger meaning of this launch is that DeepSeek has shown it can deliver high-level AI without relying on Western hardware.

By making its models work natively on multiple Chinese chips at the same time, it lowers the risk from export restrictions that have blocked Chinese companies from accessing the most powerful American processors.

Cost is also important. DeepSeek has worked hard to keep the price of running its models low. This makes it easier for businesses to build automated systems without facing very high computing costs.

In this way, Deepseek’s upgrades and launch are not mainly about one technical breakthrough. They are more about an entire supply chain coming together.

From this release, the question of who leads in AI appears to be moving away from who builds the smartest model, and toward who can keep the whole system running cheaply and independently for the long term.

If you're reading this, you’re already ahead. Stay there with our newsletter.