DeepSeek Adds Vision Mode as Chinese Chips Match Launch Support

DeepSeek, a Hangzhou-based artificial intelligence company, has added image and video recognition capabilities to its main chatbot through a new “image recognition mode” feature, bringing it in line with other major AI chatbots. The launch coincided with a significant milestone in China’s chip supply chain, as four domestic semiconductor companies—Huawei Ascend, Cambricon, Hygon Information, and Moore Threads—confirmed same-day support for DeepSeek’s newest flagship model, DeepSeek-V4, marking a shift from the previous months-long adaptation periods typical outside Nvidia’s ecosystem.

DeepSeek’s Vision and Model Expansion

DeepSeek quietly rolled out the image recognition mode alongside two other modes launched earlier in the month: “expert” and “flash.” According to Chen Xiaokang, who heads DeepSeek’s multimodal team, the tool was first tested with a small group of users on both the website and mobile app. Chen Deli, a senior researcher at the company, celebrated the launch with a post referencing the company’s logo: “The little whale can now see.”

The image and video feature arrived just days after DeepSeek released a preview of DeepSeek-V4 and made the model weights available for public download and use. V4 is structured as two distinct models: DeepSeek-V4-Pro, featuring 1.6 trillion parameters designed for complex reasoning and multi-step automated workflows, and DeepSeek-V4-Flash, optimized for handling large request volumes at lower cost. Both models support a context window of one million tokens and employ a hybrid attention design that the company says reduces computing power and memory requirements during inference.

Chinese Chip Makers Achieve Same-Day Launch Support

What captured industry attention was not merely the model itself, but the coordinated hardware support demonstrated on the day of V4’s release. Huawei Ascend confirmed compatibility with its A2, A3, and 950 chips, with the Ascend 950 using fused computing processes and parallel processing streams to accelerate inference for both V4-Pro and V4-Flash. Cambricon completed its adaptation using the open-source vLLM inference framework and published its code on GitHub. Hygon Information conducted deep model optimization on its DCU platform to enable smooth transition from model release to deployment. Moore Threads partnered with the Beijing Academy of Artificial Intelligence to run V4 on its MTT S5000 card using the FlagOS software stack.

This same-day support across multiple chipsets represents a departure from historical patterns. Previously, hardware outside Nvidia’s ecosystem typically required months to support major new models. Industry observers note that achieving compatibility across four different domestic chipsets on launch day signals a genuine shift in China’s semiconductor and AI infrastructure maturity.

Strategic Implications: Cost and Supply Chain Independence

The broader significance of DeepSeek’s launch extends beyond individual technical achievements. By enabling V4 to run natively on multiple Chinese chips simultaneously, DeepSeek reduces dependency risk from export restrictions that have historically blocked Chinese companies from accessing the most advanced American processors. Cost efficiency remains central to DeepSeek’s strategy—the company has prioritized keeping model operation costs low, enabling businesses to build automated systems without prohibitive computing expenses.

Industry observers characterize this release as reflective of an entire supply chain maturing rather than a single technological breakthrough. The coordination between DeepSeek, chip manufacturers, and software frameworks demonstrates integrated ecosystem development. This trajectory suggests the competitive landscape in AI is shifting from a focus on individual model sophistication toward capability to maintain complete, cost-effective, and independent systems over the long term.

FAQ

What new capabilities did DeepSeek add to its chatbot? DeepSeek added an “image recognition mode” that allows its chatbot to understand photos and videos, not just text. This feature was first tested with a small group of users on both the website and mobile app, bringing DeepSeek’s capabilities in line with other major AI chatbots offering similar abilities.

Which Chinese chip companies supported DeepSeek-V4 on launch day? Four Chinese chip companies confirmed same-day support for DeepSeek-V4: Huawei Ascend (with A2, A3, and 950 chips), Cambricon, Hygon Information, and Moore Threads. This same-day compatibility across multiple chipsets was previously rare outside Nvidia’s ecosystem, typically requiring months of adaptation work.

What are the two versions of DeepSeek-V4 and how do they differ? DeepSeek-V4-Pro features 1.6 trillion parameters and is designed for complex reasoning and multi-step automated workflows, while DeepSeek-V4-Flash is optimized for handling large volumes of requests at lower cost. Both support a context window of one million tokens and use a hybrid attention design to reduce computing power and memory requirements.