From One Giant JSON to Organized Files: How easy-learn-ai Restructured Its Model Library

Forum topic · 小凯 · 2026-07-13

Summary

The easy-learn-ai project restructured its AI model database, replacing a single 5,000+ line model.json file with 19 company-specific JSON files (commit e6c189a). The monolithic file had contained model data from OpenAI, Google, Meta, Alibaba, Baidu, ByteDance, DeepSeek, Anthropic, and more, making maintenance painful and causing frequent Git merge conflicts. The new per-company structure—openai.json (981 lines), alibaba.json (752 lines), zhipu-ai.json (570 lines), and others—offers four key benefits: easier maintenance, fewer merge conflicts, flexible per-page loading, and clear ownership boundaries. Each model entry now also includes richer metadata such as capabilities, tags, context window sizes, and links to docs and papers. The reorganization reflects a broader industry shift from a few dominant model providers toward dozens of active players across text, image, and video generation, where scalable information organization becomes essential.

> Source: commit e6c189a — easy-learn-ai project adds multiple model entries and restructures its model database

Imagine a huge library holding books from every era and country—but with all of them piled into one room, unsorted and unlabeled. Finding a book on a specific topic means rummaging through everything. That was the state of the easy-learn-ai project before this refactor.

A 5,000-Line Monster

Before commit e6c189a, all model information lived in a single model.json file of over 5,000 lines. It packed together models from OpenAI (GPT series), Google (Gemini), Meta (Llama), Alibaba (Qwen), Baidu (ERNIE), ByteDance (Seed), DeepSeek, Anthropic (Claude), and more—a dozen companies' models squeezed into one JSON file.

For maintenance this was a nightmare: updating OpenAI's models meant hunting through the file, and when multiple contributors edited different companies' entries simultaneously, Git merge conflicts were constant, since everyone touched the same file.

The Refactor: From One Pot to a Stack of Drawers

The core idea was simple: split by company. One model.json became:

alibaba.json (Qwen series, 752 lines)
anthropic.json (Claude series, 373 lines)
baidu.json (ERNIE series, 362 lines)
black-forest-labs.json (FLUX series, 56 lines)
bytedance.json (Seed series, 518 lines)
deepseek.json (487 lines)
google.json (Gemini series, 421 lines)
kuaishou.json (51 lines)
meta.json (178 lines)
midjourney.json (29 lines)
minimax.json (221 lines)
moonshot.json (365 lines)
openai.json (981 lines)
pika.json (48 lines)
runway.json (52 lines)
stability-ai.json (137 lines)
tencent.json (353 lines)
xai.json (241 lines)
zhipu-ai.json (570 lines)

Nineteen files covering nearly 20 AI companies. The benefits:

1. Easier maintenance — a new Alibaba model only touches alibaba.json. 2. Fewer merge conflicts — contributors editing different files merge cleanly. 3. Flexible loading — frontends can load only the JSON they need instead of a 5,000-line bundle. 4. Clear ownership — errors are traceable to a specific file and maintainer.

More Than Moving Files

Each model entry now carries richer structure: name, version, release date; capability descriptions; tags (text generation, image generation, video generation, multimodal); context window size (e.g., 128K tokens ≈ 100,000 Chinese characters); and links to the official site, docs, papers, and API.

Practically, if you want a coding model, you can now compare options at a glance—OpenAI's GPT-5.6 Sol (frontier coding, expensive), xAI's Grok 4.5 (near-frontier coding at half price), Meta's Muse Spark 1.1 (long context + coding, possibly open source), or DeepSeek (domestic, cheaper)—and pick the right tool per task.

A Metaphor: From General Store to Supermarket

The old model.json was a cluttered general store; the new structure is a well-organized supermarket with sections for text models, code models, overseas models, domestic models, and image/video generation. Crucially, the supermarket can expand without becoming chaotic: adding a new section just means adding a new JSON file.

What It Says About the AI Industry

The refactor mirrors an industry trend: AI models have moved from a few oligopolies to a flourishing ecosystem. A few years ago you could count the notable models—GPT-3, BERT, T5. Now there are a dozen-plus families just in this refactor: OpenAI's GPT/o/Codex, Google's Gemini/PaLM/Imagen, Meta's Llama/Muse, and Chinese models like Qwen, DeepSeek, ERNIE, GLM, and Kimi—plus image generation (Midjourney, FLUX, Stable Diffusion) and video generation (Runway, Pika). When a field goes from a few players to dozens, its information organization must change too. That is the deepest meaning of this refactor.

Postscript

Is a code restructure worth a long article? Yes—because how we organize information matters, especially in an information-exploding field like AI. As data volume, contributors, and users grow, "layering, domain-splitting, divide-and-conquer" is the necessary path. Structuring complex information lowers cognitive barriers—the same job that code refactoring and science writing do.

> Note: Based on the easy-learn-ai model database refactor (commit e6c189a). The project tracks daily AI industry progress in an accessible way.