DeepSeek Launches Image Recognition but Can’t Identify Its Own Founder
Chinese AI startup DeepSeek has officially launched an image recognition mode on its web platform, marking its transition from a pure text-based model to a multimodal AI system. The rollout comes just days after the company confirmed it had secured over 50 billion yuan (~$7 billion USD) in its first-ever external funding round — the largest single fundraising in Chinese AI history. However, in a widely-reported quirk that has captured the internet’s attention, the AI failed to recognize its own founder, Liang Wenfeng, during testing.
Context
DeepSeek, backed by Liang Wenfeng’s High-Flyer quant fund, had long been known for its staunchly anti-commercialization stance and refusal to take outside capital. That changed in mid-June 2026, when the company completed a landmark 51 billion yuan funding round at a valuation estimated between 330 and 400 billion yuan (~$48-58 billion). According to The Paper, Liang Wenfeng personally contributed approximately 20 billion yuan as the largest single investor, with Tencent investing 10 billion yuan and CATL, NetEase, JD.com, and several venture capital firms also participating.
Key Developments
Image Recognition Launch
On June 18, DeepSeek multimodal researcher Xiaokang Chen announced the full rollout of the image recognition feature, which had been in gray-scale testing since late April. The technology is built on DeepSeek’s proprietary DeepSeek-OCR2 with a visual causal flow mechanism, enabling object recognition, scene parsing, chart analysis, and text extraction. The mobile app version remains in internal testing.
The Founder Recognition Failure
When journalists and users tested the new feature with a photo of Liang Wenfeng, the results were unexpected. After several minutes of “thinking,” the AI generated multiple incorrect identifications, misidentifying the DeepSeek founder as Moonshot AI founder Yang Zhilin, a “younger version” of Tencent founder Ma Huateng, Meituan founder Wang Xing, and even Xiaomi CEO Lei Jun. Ultimately, the AI admitted: “I’m really not sure. I don’t recommend guessing names, because guessing wrong is worse than saying ‘I don’t know.’”
Independent testing by IT之家 confirmed the results, noting that while DeepSeek could not recognize Liang Wenfeng, it accurately identified multiple photos of Lei Jun. The likely explanation is that Liang Wenfeng is known to be extremely low-profile with limited public photos available online, making it difficult for the model to form stable recognition features. The failure also suggests DeepSeek did not specially optimize recognition for its own founder.
Broader Testing Results
Extensive testing by Machine Heart Pro revealed a mixed picture of the new feature’s capabilities. The model successfully identified Nvidia CEO Jensen Huang but struggled with basic tasks: it misidentified “douzhi” (soybean milk) as milk, failed 4 out of 7 handwriting recognition tests, could not match identical socks in a puzzle, and incorrectly identified a piano chord. Notably, competitors Gemini 3.5 Flash, GPT 5.5, and Claude Sonnet 4.6 also failed the piano chord test, suggesting this remains a challenge across the industry.
Analysis
The launch represents a significant milestone for DeepSeek, filling what industry observers described as its “last shortcoming” — the absence of multimodal capabilities. However, the feature is clearly in early development, with notable accuracy issues in face recognition, handwriting, and specialized tasks.
More significantly, the timing of the launch — just days after the record funding round — signals DeepSeek’s strategic shift. The company, which previously rejected all venture capital overtures, has now accepted that massive capital is necessary to compete in the global AI arms race. The investment structure is notably founder-friendly: external funds flow through limited partnerships managed by Liang Wenfeng, ensuring his absolute control with a five-year lock-up period for investors. According to The Paper, Liang Wenfeng’s primary requirement for all investors was a strict “no poaching” condition — forbidding them from recruiting DeepSeek employees or suggesting they start competing ventures.
What’s Next
The image recognition mode’s current limitations raise questions about when DeepSeek will release a developer API for multimodal capabilities and whether technical documentation on the underlying architecture will follow. With 50 billion yuan in fresh capital, the company is well-positioned to accelerate R&D, particularly in compute infrastructure — it has already begun hiring for GW-level data center construction. As DeepSeek continues its stated mission of “moving forward steadily in experimentation and reflection, striving to get closer to the goal of achieving AGI,” the coming months will reveal whether this quirky founder-recognition failure becomes a footnote or a catalyst for rapid improvement.