When AI Models Get Their Own Household Registry: A Gentle Revolution in Knowledge Organization

Have you ever wondered how the model information you see on AI learning sites—GPT-4, Claude, ERNIE Bot—gets organized and stored?

The answer: it used to be a hodgepodge, all crammed into one file.

I. From "One-Pot Stew" to Separate Registries

Imagine walking into a giant library where one room holds a single, enormously thick catalog. It records everything from *The Analects* to *Harry Potter*—title, author, publisher, page count, recommended age—all mixed together, sorted alphabetically.

If you wanted "Chinese science fiction," you'd flip through the entire book. If you wanted "all books from a certain publisher," you'd flip through it again.

That was the previous state of the easy-learn-ai project: all model info—Alibaba's Qwen, OpenAI's GPT series, Baidu's ERNIE, Google's Gemini—crammed into one giant model.json file with over 5,000 lines, like an overstuffed suitcase.

This commit did something simple yet crucial: it gave each vendor its own household registry. Alibaba's models go into alibaba.json, Anthropic's into anthropic.json, Baidu's into baidu.json, ByteDance's into bytedance.json... 18 new vendor files in total, each recording only its own children.

II. Why Split the Files?

You might ask: isn't one file more convenient? Actually no. Think of a teacher with 50 students: instead of one huge folder, records are sorted into drawers by gender and height—search efficiency improves by orders of magnitude.

In software engineering this is called modularity or domain separation. Its benefits include:

1. Each parent handles its own child. Baidu updates ERNIE 4.0 parameters? Just edit baidu.json without risking Google's Gemini config. This is reducing coupling—a change in one place doesn't cascade.

2. A clear map at a glance. The project directory now reads like a vendor roster:

• alibaba.json — Qwen, China's open-source pride
• anthropic.json — Claude, the AI that writes poetry
• bytedance.json — Doubao and the Seed series
• deepseek.json — the reasoning king R1
• google.json — Gemini, the multimodal ambition
• openai.json — the GPT family, the industry benchmark

Like switching from a blurry satellite image to a high-resolution administrative map.

3. Parallel collaboration becomes possible. Two people editing model.json simultaneously would inevitably conflict—like two people stuffing one suitcase. Now each contributor edits their own files.

4. Future-proof extensibility. A new AI company appears tomorrow? Create newvendor.json, fill in the info, and the project recognizes it automatically—no touching the 5,000-line monster.

III. What Secrets Hide in the New Data?

Each file follows a carefully designed schema:

modelName, company, country, openSourceStatus, releaseDate, description, modelTags, contextWindow, maxGenerationTokenLength, relatedLinks

Each field answers a user question:

• country: Where is this model from? Chinese users care about domestic progress.
• openSourceStatus: Can I run it locally? Core for developers and privacy-sensitive users.
• contextWindow: How much can it "remember" at once? 64K means it can read a novella and discuss the plot.
• maxGenerationTokenLength: How long can one answer be? Enough for a full weekly report?
• relatedLinks: Papers, GitHub repos, API docs for deeper exploration.

This structure turns raw data into knowledge assets that machines can parse, humans can read, and programs can consume.

IV. DeepSeek's Distillation Family: A Case Worth Savoring

deepseek.json is especially interesting. It records not just flagship DeepSeek-R1 but also its "children"—a series of distillation models.

Distillation is like a martial arts master passing their lifetime of skill to several apprentices: the weakest learns thirty percent but is light and agile; the strongest learns seventy to eighty percent and can stand on their own.

DeepSeek-R1-Distill-Qwen-1.5B / 7B / 14B / 32B, plus Llama-based 8B and 70B—same father, different children, suited to different contexts:

• 1.5B: lightweight local deployment, teaching practice—a small scooter, fine for short trips
• 32B: near or better than o1-mini on multiple reasoning benchmarks—approaching top-tier performance
• 70B: hard reasoning, complex code generation, production-grade agents—the enterprise heavy artillery

Crucially, all of these models are open source. Anyone can download, use, modify, even commercialize them without paying DeepSeek a cent. In an increasingly closed industry, that openness is itself a statement.

V. The Bigger Picture

Viewed together, the 18 files sketch an emerging map of the AI world:

• Chinese camp: Alibaba (Qwen), Baidu (ERNIE), ByteDance (Doubao/Seed), Moonshot (Kimi), Zhipu AI (GLM), DeepSeek, Tencent, Kuaishou, MiniMax...
• American camp: OpenAI (GPT), Anthropic (Claude), Google (Gemini), Meta (Llama), xAI (Grok), Midjourney, Runway, Pika, Stability AI...

This is not just a contest between two countries but a collision of two philosophies: Chinese vendors are catching up at astonishing speed, in some areas (DeepSeek's reasoning, Qwen's multilingual capability) already in the world's top tier; American vendors still hold the high ground in compute and talent, but open source is loosening their monopoly.

This update is essentially drawing the map. In a gold rush, the safest business isn't gold—it's selling shovels. In the AI arms race, the most underrated value isn't training a model; it's helping people see the whole board and understand each model's position and meaning.

VI. Thoughts Beyond Technology

When a field explodes with growth, you face two choices: pile everything together and pray you don't get lost, or deliberately design a classification system and accept that "classification is itself thinking."

easy-learn-ai chose the latter. With 18 files it answers a meta-question: in the AI world, "vendor" is a more stable classification dimension than "function," because functions change (today's text model generates video tomorrow) while vendor boundaries are relatively clear. That choice is itself architectural wisdom.

Epilogue

Next time you open easy-learn-ai and see those neatly arranged model cards, remember this story: someone faced a 5,000-line monster and decided to take it apart, issuing each model its own registry. He trained no new model and published no paper, but he did something equally important—he made knowledge visible, accessible, and understandable. In that sense, he too is an architect. What he built is not a model, but a bridge to models.

*This article is based on easy-learn-ai commit e6c189a.*