DeepSeek Finally Got Eyes — And It Can't Even Recognize Its Own Boss

Let me tell you something funny.

DeepSeek — the AI company that's been flexing its text muscles for over a year — finally added vision to its toolkit. You can now upload a photo and have it actually see what's in it. Sounds cool, right?

Here's the punchline: Show it a picture of its own CEO, Liang Wenfeng, and it'll confidently tell you it's Jack Ma. Or Pony Ma. Or some random guy. It has no freaking clue who its own boss is [citation:1].

This went live about 15 hours ago [citation:2]. And honestly? It's exactly the kind of chaotic energy I've come to expect from DeepSeek. They nail the engineering, but the polish? Not quite there yet.

DeepSeek's new Vision Mode gives AI the ability to "see" — but the execution is uneven. Image via Unsplash

Okay, But What Does This Thing Actually Do?

First things first: this isn't just a glorified OCR tool. It's a full visual understanding system that sits right next to "Fast Mode" and "Expert Mode" in the DeepSeek interface [citation:3].

The technical magic behind it is called "Thinking with Visual Primitives" [citation:4]. Here's the dumbed-down version:

This matters because most vision models are terrible at "pointing." They'll say "the thing on the left" and you have no idea what they mean. DeepSeek's approach is different — it anchors its reasoning to actual coordinates in the image [citation:5].

The efficiency is also wild. Processing an 800×800 image costs roughly 90 tokens — while Claude Sonnet needs about 870 and Gemini needs around 1,100 [citation:6]. That's not just clever. It's competitive advantage territory.

DeepSeek's "visual primitives" approach points at objects with coordinates, not just words. Image via Unsplash

The Good, The Bad, and The Straight-Up Embarrassing

✅ What It Nails

When it works, it really works.

Test it with a Shanghai skyline photo — it IDs four major buildings and correctly calls out "the white bridge is probably Zhapu Road Bridge," which is a classic photography spot [citation:7].

Show it a museum artifact, and it'll describe textures and infer historical styles. Show it a meme, and it gets the joke [citation:8]. Show it a webpage layout, and it can spit back working code [citation:9].

For screenshots, tables, documents, and flowcharts, it's fast and reliable [citation:10]. That's genuinely useful for anyone who works with messy digital content.

❌ Where It Faceplants

But here's the problem: the failures are spectacular.

It can't recognize its own CEO. Multiple tests show DeepSeek misidentifying Liang Wenfeng as Jack Ma, Pony Ma, or just saying "I really don't know" after thinking for minutes [citation:11].

It doesn't know anyone recent. Show it Cape Verde goalkeeper Vozinha — who went viral in the last week — and it'll "think" for over a minute, mention Cape Verde multiple times, then give a completely wrong answer [citation:12].

Why? Because the training data stops at 2025 [citation:13]. And there's no internet search in Vision Mode — unlike competitors like Doubao that auto-connect to search results [citation:14].

🤡 The Embarrassing Part

Here's where it gets really awkward. In late April, DeepSeek published their technical report on this vision model — then took it down hours later. The GitHub page returned a 404 [citation:15].

Was it leaked? Was it incomplete? Did they reveal too much? Nobody knows. But it's not a good look for a company that just raised a reported $7 billion in Series A funding [citation:16].

Why I Still Care About This (And You Should Too)

Look, it's easy to clown on DeepSeek for the CEO fail. But this isn't really about celebrity recognition.

This is about agents.

DeepSeek's V4 model, released in April, was all about Agent capabilities — getting AI to actually do things. But a text-only agent hits a ceiling fast. If you want AI that can:

• Navigate browser interfaces
• Read dashboards
• Interpret UI elements
• Analyze complex charts and PDFs

…it needs to see. Full stop [citation:17].

And DeepSeek might be the late but smart player here. While others rushed to slap on mediocre multi-modal features, DeepSeek waited until they had something different — something that prioritizes pointing accuracy over pixel resolution [citation:18].

The early benchmarks back this up. In counting tasks, DeepSeek scored 89.2% vs Gemini's 88.2% and GPT-5.4's 76.6%. In maze navigation, it scored 66.9% vs GPT-5.4's 50.6% [citation:19]. That's not just competitive. That's dominant in specific niches.

DeepSeek's vision mode is a prerequisite for true AI agents — and they might be late, but they're not wrong. Image via Unsplash

The Bottom Line

DeepSeek finally gave its AI eyes — and it's already embarrassing itself in public.

But honestly? That's kind of what makes this fun to watch. AI is still figuring out how to point at things without looking like a toddler. And DeepSeek, for all its flaws, is trying something genuinely different than the pixel-stacking approach everyone else is taking.

The CEO thing? It's funny. It's embarrassing. It'll get fixed.

The deeper question is: does this approach actually work better in the long run? My gut says yes. But I'm not holding my breath until I see it handle a real-world dashboard without hallucinating half the data.

What's your experience been with DeepSeek's Vision Mode? Drop a comment below — I'm genuinely curious.