Background
Imagine a huge library where every book from every era is piled into one room with no labels or categories. That was the state of the easy-learn-ai project before commit e6c189a: all AI model information lived in a single model.json file of over 5,000 lines.
That one file contained model data from more than a dozen companies: OpenAI's GPT series, Google's Gemini, Meta's Llama, Alibaba's Qwen, Baidu's ERNIE, ByteDance's Seed, DeepSeek, Anthropic's Claude, and more.
The Problem with a 5,000-Line Monolith
- Hard to maintain: adding a new OpenAI model meant digging through thousands of lines to find the right section.
- Frequent merge conflicts: when multiple contributors edited different companies' entries in the same file, Git could not cleanly merge their changes, causing constant conflict resolution.
- Basics: name, version, release date
- Capabilities: what the model is good at and suitable for
- Tags: text generation, image generation, video generation, multimodal, etc.
- Context window: how much dialogue the model can retain (e.g., 128K tokens ≈ 100,000 Chinese characters)
- Links: official site, docs, papers, API entry points
The Refactor: One Drawer Per Company
The fix was simple: split by company. The single model.json became 19 files:
| File | Lines | File | Lines |
|---|---|---|---|
| openai.json | 981 | tencent.json | 353 |
| alibaba.json | 752 | anthropic.json | 373 |
| zhipu-ai.json | 570 | baidu.json | 362 |
| bytedance.json | 518 | moonshot.json | 365 |
| deepseek.json | 487 | minimax.json | 221 |
| google.json | 421 | meta.json | 178 |
| xai.json | 241 | stability-ai.json | 137 |
| black-forest-labs.json | 56 | runway.json | 52 |
| kuaishou.json | 51 | pika.json | 48 |
| midjourney.json | 29 | | |
The benefits:
1. Easier maintenance — updating Alibaba's models only touches alibaba.json; Google updates only touch google.json.
2. Fewer merge conflicts — two people editing different companies now change different files, so Git merges automatically.
3. Flexible loading — a page showing one company's models can load just that JSON instead of the full 5,000-line bundle, improving page speed.
4. Clear ownership — errors are easy to trace to a specific file.
Not Just a Move, an Upgrade
Each model entry now carries richer structured information:
An Industry in Miniature
The refactor mirrors a broader trend: AI has moved from a handful of dominant models (GPT-3, BERT, T5) to dozens of families across text (GPT, Gemini, Llama, Qwen, DeepSeek, ERNIE, GLM, Kimi), image (Midjourney, FLUX, Stable Diffusion), and video (Runway, Pika). When a field goes from a few players to dozens, the way information is organized must change too.
Takeaway
When data volume, contributors, and users all grow, "layering, domain-splitting, divide-and-conquer" becomes inevitable. Structuring complex information — whether via code refactoring or explanatory writing — lowers the cognitive barrier and makes knowledge more accessible. Teams running similar AI tracking projects can take this per-vendor, per-file structure as a reference.
> Based on the easy-learn-ai model database refactor, commit e6c189a. The project tracks daily AI industry progress in plain language.