Have you ever tried flipping through a really thick phone book?
As a child, I saw one of those yellow pages at my grandmother's house — a huge thick volume, with every business's name, phone number, and address crammed together in pinyin order. Looking for a restaurant? First find the "dining" category under "C", then slowly scan through hundreds of names. It felt like searching for a needle in a haystack: the information was all there, but between you and what you actually needed stood a wall.
The AI industry today faces a similar problem — except behind that wall are not phone numbers, but the names, parameters, capabilities, and release dates of over two hundred AI models.
---
An Ever-Thickening "Phone Book"
In early 2023, if you wanted to understand the large language models on the market, you probably only needed to remember one name: GPT-4. The AI circle was small back then, models were few and far between — like a small town's directory that fit on a single page.
Two years later, the situation is completely different.
In the latest commit to the easy-learn-ai project, the developers did something that looks tedious but is deeply meaningful: they split the model data that had been crammed into one giant file into 19 independent files, one per company — Alibaba, Anthropic, Baidu, Black Forest Labs, ByteDance, DeepSeek, Google, Kuaishou, Meta, Midjourney, MiniMax, Moonshot AI, OpenAI, Pika, Runway, Stability AI, Tencent, xAI, and Zhipu AI.
How many models in total? 243.
What does this number mean? If you learned one new model per day, it would take you most of a year to know them all. And these are only the "mainstream" models included in this project — not counting the vast numbers of small models, vertical-domain models, and enterprise-internal models.
The AI industry has gone from "a table with a few players" to a bustling shopping mall.
---
From One Mixed Pot to Categorized Shelves
So why is splitting files worth talking about?
Because the way knowledge is organized determines how deeply you can understand it.
Imagine walking into a bookstore. If all the books were piled in one big room, randomly ordered by arrival time, could you find the book you want? Technically yes — if you were willing to spend enough time. But good bookstores have sections: literature, history, science, art... with sub-sections within each. This isn't for looks; it's because the human brain is inherently good at processing hierarchical, structured information.
The restructuring of the AI model database follows the same principle.
Cramming all models into a single model.json was like dumping all books into one basket. Want to compare Alibaba's and Baidu's model development trajectories? You'd have to manually filter through a jumble of mixed data. Want to see which company is most deeply invested in video generation? There was no intuitive way.
Now, with one file per company — like a dedicated shelf per publisher in a bookstore — you can see at a glance:
- OpenAI has 38 models, the most cataloged — from early GPT-3.5 to the future GPT-5.5 Pro, it has clearly kept iterating;
- Alibaba has 28, with the Qwen series evolving from 1.0 to 3.7-Max, pursuing a dual "open-source + closed-source" strategy;
- DeepSeek, with "only" 17 models, proved that quality can beat quantity when R1 and its distilled versions caused a sensation in early 2025;
- ByteDance has 20, from the early Cloud Skylark model to doubao-seed-2.0-pro, showing the company's aggressive AI investment;
- while Black Forest Labs (creators of FLUX) has just 2 models — yet enough to shake up the image generation field.
- DeepSeek-R1 stunned the world in early 2025 with an open-source + ultra-low-price strategy, proving "small teams can build big models";
- The Qwen series has kept iterating through 3.5, 3.6, 3.7, and its influence in the open-source community keeps growing;
- Kimi is known for long context windows, among the first to put 2-million-character context into practical use;
- ByteDance iterates its models extremely fast, deeply integrated with its content ecosystem (Douyin, CapCut).
- Where are your users mainly located? (affects latency and compliance)
- Is your task chatting, coding, image generation, or video editing? (different models excel at different things)
- What's your budget? (from completely free open-source models to expensive flagship APIs, prices can differ a hundredfold)
- How much context do you need? (from 4K to 1M — needs vary enormously across scenarios)
- How sensitive are you to data privacy? (some models can be deployed locally, others must be called online)
This structure lets you, for the first time, make horizontal comparisons — not of who has more parameters, but of the paths each company has chosen: some spread wide, some go deep; some open-source, some closed; some all-in on text, some simultaneously building image and video capabilities.
---
A Map, Not a Pile of Bricks
The most interesting part of this update isn't just the swelling numbers, but the diversification of model types.
Look closely at those 19 companies and you'll see the battlefields now extend far beyond the original "chatbot" track:
Pure text LLMs remain the most crowded track — Qwen, Claude, GPT, Gemini, DeepSeek, Kimi, GLM... names you probably already know well. They compete on reasoning ability, context length, coding ability, and Agent execution.
Right behind comes image generation: Black Forest Labs' FLUX series, Midjourney, Stability AI's Stable Diffusion family, ByteDance's SeedEdit... These models don't converse — they paint. You type a sentence, they output an image. This track has been exploding since Stable Diffusion went open-source in 2022, and now has a complete "industrial system."
Then comes video generation — possibly the most exciting field of 2024–2025. Pika, Runway, Kuaishou's Kling, ByteDance's Jimeng... These companies are teaching AI to understand "time." An image is static, but video flows; drawing a single frame isn't hard — the hard part is making 24 consecutive frames form a coherent motion, a physically plausible world. This track's technical barriers are higher than image generation's, but the imaginative space is bigger too.
What does this tell us?
AI is no longer a "thing" — it's an "ecosystem." Just as when electricity was first invented people assumed its only use was lighting, but it soon drove motors, heated ovens, ran trams, and connected the world, AI is going through the same diffusion — from "a program that can chat" to an all-purpose assistant that "can see, draw, edit video, write code, and control computers."
---
The Chinese Players' "Group Army"
Looking closely at those 19 companies, you'll notice something interesting: Chinese companies make up nearly half.
Alibaba (Qwen), Baidu (ERNIE), ByteDance (doubao), DeepSeek, Moonshot AI (Kimi), Zhipu (GLM), MiniMax, Tencent (Hunyuan), Kuaishou (Kling) — 9 Chinese companies, each with its own flagship models and unique路线... each with its own flagship models and distinct路线.
This was unthinkable in the AI landscape of two years ago. In early 2023, when large models came up, most people only knew OpenAI and Google. Chinese companies were mostly "followers" — releasing GPT-benchmarked products, but always half a beat behind in technology and ecosystem.
Today, that picture has been completely rewritten:
These 9 Chinese companies together account for over 100 cataloged models. If AI is a war, China is no longer a "guerrilla force" but a regular army with a complete "group army structure."
---
Why Does This Matter to You?
By now you might ask: what do model counts and file splits have to do with me?
The connection is: choice matters more than effort — and the precondition for choosing is knowing what the options are.
Suppose you're a developer who wants to integrate an AI model into your application. Two years ago this was simple — just use OpenAI's API. Today, you need to consider:
---
Conclusion: The Shape of Knowledge Determines the Quality of Thinking
One easily-overlooked change in this commit: src/utils/modelApi.ts was modified by 55 lines. What is it for? It's the "bridge" connecting the frontend display to the backend data — when the file structure changed, the data-reading code had to change with it.
This is really a microcosm of the whole story: when the way knowledge is organized evolves, the way we use knowledge must evolve too.
243 models, 19 companies, 3 media types (text/image/video) — this isn't just a numbers game. It's a sign of the AI industry moving from "infancy" to "adolescence" — the body growing fast, organs differentiating, capabilities diversifying. And we observers, users, and participants need to learn new ways to understand this rapidly expanding world.
The library has been built. The shelves are organized. Now it's our turn to read the books.
---
*Commit: e6c189a | easy-learn-ai model database restructuring and expansion* *Models cataloged: 243 | Companies covered: 19 | Data span: 2023–2026*