380 Million Users Can’t Be Wrong… Except They Are

Here’s a number that should make you question everything: 3.82 billion monthly active users. That’s not a small country’s population. That’s Doubao, the AI assistant that supposedly rules China. And here’s the number that should make you laugh out loud: that 3.82 billion-user giant generates less than $1 million in revenue a day. That’s less than 0.003 yuan per user. Per day.

Let that sink in.

You have an app that everyone has downloaded, but nobody is paying for. And on the other end of the spectrum, you have Kimi—the darling of academics, the ‘paper-reading wizard’—with a paltry 7.29 million monthly users. Less than one-fiftieth of Doubao’s base. Yet its users are fiercely loyal.

This is the great AI split: You have one product winning the popularity contest, another winning the profitability race, and a third winning the respect of the people who actually use it. And if you’re a student trying to figure out which one to keep on your phone, the marketing noise is deafening.

So I did something radical. I ignored the marketing. I spent a month using all three as a student, for real tasks—writing outlines, debugging code, and synthesizing information. Here’s what I found.

The Brutal Truth: No One Is Winning Everything

Let’s start with the numbers, because they tell a story that’s both fascinating and tragic.

Kimi’s problem isn’t that it’s bad. It’s that its core strength—reading long documents—is a niche, low-frequency need. Nobody reads a paper every day. That’s why its user retention time is nearly identical to Doubao’s (26.1% vs 27.5%), but its total user base is tiny. The product is sticky; it just doesn’t attract a crowd.

DeepSeek, on the other hand, has cracked the code. Its ‘search + reasoning’ positioning captures real, daily needs like checking facts or writing code. That’s why it has a solid 130 million users and the highest retention in the game. It’s not a toy. It’s a tool.

The 30-Day Test: Three Personalities, Three Fatal Flaws

I threw three high-frequency student tasks at these apps to see which one wouldn’t just survive, but actually help.

Test 1: Creating a Structured Study Outline

I asked them to turn a chunk of a social psychology textbook into a markdown study guide with a comparison table.

DeepSeek was the efficiency king. In 39 seconds, it delivered a concise, logically-perfect outline. The table didn’t just list differences; it included a column for ‘interconnections,’ showing a deep understanding of the material. It was print-and-memorize ready.

Doubao took 1 minute 47 seconds and gave me a novel. It was comprehensive but bloated, burying the key points in a wall of text. I’d need to spend another hour just to condense it.

Kimi took 4 minutes and 10 seconds—six times slower than DeepSeek—but returned a masterpiece. It included memory tricks, common exam angles, and a structure so clear it could be a published study guide. But the wait time is a killer when you’re studying at 2 AM.

Test 2: The Multi-Turn Editing Loop

This is where things get interesting. I asked each AI to write an outline for a paper, then refine it, then completely re-style it as a policy brief.

DeepSeek was the undisputed champion. It nailed the context, remembered the earlier requirements, and executed the restructuring perfectly. It even suggested concrete policy proposals with actionable details.

Doubao was stubborn. Its first draft was okay, but when I asked for a deep restructure, it just sprinkled new words on the old skeleton. It’s fine for a grammar check, terrible for a revision.

Kimi had a moment of glory in the second turn, adding excellent simulated data. But in the third turn, it crumbled. It lost the thread, merged the detailed sections I’d just approved into a generic blob, and ignored part of my instructions. It’s like a brilliant friend who gets distracted halfway through a conversation.

Test 3: The Code Debug

I gave all three a Python script with five hidden bugs, from a fatal ZeroDivisionError to poor naming conventions.

DeepSeek found all five. It prioritized them perfectly, explained the crashes first, then the code hygiene issues. It even added an extra check for an empty list—a level of engineering paranoia that’s rare in AI.

Doubao fixed the obvious errors but missed the fatal divide-by-zero bug. Even worse, it “fixed” a floating-point issue by rounding the average down to an integer, losing precision. It solved one problem by creating another.

Kimi correctly identified the critical crashes but handed back a fix with broken indentation. Copy-paste it, and Python screams at you. For a coding task, that’s an automatic fail.

The Ugly Side of ‘Smooth’

There’s a reason Doubao has 3.82 billion users: it’s the most polished. Voice input, photo questions, it feels like a modern consumer app. But that polish comes at a cost. It’s designed to give you the answer that feels best, not the answer that is correct. It’s a people-pleaser, not a truth-teller.

DeepSeek’s interface feels like a relic, but it has a giant, reassuring ‘Deep Think’ button. It’s a statement of confidence: “You’re here for the reasoning. We’ll handle the rest.”

Kimi’s fatal flaw is that its interface doesn’t support its own superpower. The process of uploading, parsing, and querying long documents is riddled with friction. It’s a Ferrari with bicycle tires.

The Bizarre Business of AI

Here’s the most absurd part of this entire saga.

Doubao’s 3.82 billion users can’t even cover the cost of their own compute power. That’s why they introduced a subscription plan costing up to $700 a year. It’s an act of desperation masked as a premium offering.

DeepSeek is playing a different game. It’s free for consumers, but it’s charging developers who rely on its API. They’re making an estimated $400-500 million a year and just raised their prices. Their moat is the quality of the model itself.

Kimi is stuck. It doesn’t have the user base to scale an API business, and it doesn’t have Doubao’s reach to sell subscriptions. It’s the classic case of a great product with a broken business model.

The Verdict That Will Piss You Off

After 30 days of brutal testing, my conclusion is simple: Don’t be loyal to any of them. Be loyal to the task.

• Need to understand a concept or fix code? Use DeepSeek. It’s free and its reasoning is unmatched.
• Need to digest a 50-page PDF? Use Kimi. Forget the user count; for long texts, it’s still the king.
• Need to quickly ask a question while walking? Use Doubao. The experience is genuinely great.

But this rational conclusion exposes a massive, gaping hole in the market. None of these companies are building for the student. They’re all fighting to be the ‘universal assistant,’ which means they’re experts in everything and masters of nothing.

Doubao has a side-app for learning that has 19 million users, but the main app has zero student-specific design. Kimi has the best long-text tech but hasn’t built a ‘reading-to-notes’ workflow. DeepSeek is the smartest kid in the room, but it can’t even export a decent note file.

They’re all trying to win the battle for the front door, while ignoring the rooms inside the house. The next winner won’t be the next Doubao or Kimi. It’ll be the one who finally builds an AI that lives in the student’s workflow, not just in their app drawer. And right now, that product doesn’t exist.