When the AI World Explodes, Someone Quietly Builds a Map: Restructuring a 5,000-Line Model Database

When the AI World Explodes, Someone Quietly Builds a Map

1. A 5,000-Line Catch-All That Became Unbearable

Imagine a library where novels, textbooks, recipes, manuals, and comics are all thrown into one giant cardboard box. Finding any single book means emptying the entire box. That was the situation of src/utils/model.json in the easy-learn-ai project.

Five thousand lines in a single JSON file, stuffed with everything from OpenAI's GPT series and DeepSeek's R1 family to Google's Gemini, Alibaba's Qwen, xAI's Grok, and Stability AI's image models.

Nobody lacked the desire to organize it—the problem is that organizing has become nearly impossible when AI models iterate on a weekly basis. You just learned Claude 3.5 and Claude 4 arrives; you finally grasp Qwen2.5 and the next version ships. DeepSeek's distilled models line up from 1.5B to 70B like Russian nesting dolls.

But someone did it.

2. Splitting One Tangled Mess into Nineteen Threads

The core action of this commit is a single word: split.

src/utils/model.json — deleted, 5,000 lines, gone.

In its place, nineteen tidy files under src/data/models/:

• alibaba.json — Alibaba's Qwen universe
• anthropic.json — Anthropic's Claude family
• baidu.json — Baidu's ERNIE series
• black-forest-labs.json — FLUX image generation
• bytedance.json — ByteDance's Seed and Doubao
• deepseek.json — the full DeepSeek lineup
• google.json — the Gemini saga
• Plus kuaishou.json, meta.json, midjourney.json, minimax.json, moonshot.json, openai.json, pika.json, runway.json, stability-ai.json, tencent.json, xai.json, zhipu-ai.json

Nineteen vendors, nineteen household registries. This isn't just file-level tidying—it reflects a cognitive clarity: when facing an explosively growing domain, categorization isn't optional; it's essential.

You wouldn't put toothbrushes and broccoli on the same supermarket shelf. Likewise, researchers of open-source ecosystems can open deepseek.json and meta.json; those interested in closed commercial models go straight to openai.json and anthropic.json; fans of Chinese models find alibaba.json, baidu.json, and zhipu-ai.json lined up neatly.

Classification is the first step toward understanding.

3. What a Model's ID Card Looks Like

After the split, every model received a standardized "identity card":

• modelName — e.g., "DeepSeek-R1" or "Claude Opus 4.8"
• company — who created it
• country — "China" or "USA"
• openSourceStatus — open or closed, one of the most important political boundaries in the AI world
• releaseDate — date of birth
• description — a plain-language introduction
• modelTags — capability labels: text generation, deep reasoning, visual understanding, code enhancement, tool calling
• contextWindow — how much it can read at once: 64K? 128K? 1M?
• maxGenerationTokenLength — how much it can write at once
• relatedLinks — official docs, papers, GitHub repos
• parent — its lineage, e.g., DeepSeek-R1-Distill-Qwen-7B's parent is DeepSeek-R1

A unified schema is the soul of any database. Without it you have only data fragments; with it, querying, filtering, and comparing become possible. Now you can ask: "Which open-source Chinese models exist?" Filter every file for country == "China" and openSourceStatus == "open source"—immediately you get the full DeepSeek-R1 series, Alibaba's Qwen3.5-Plus, ByteDance's Seed-OSS-36B-Base, Zhipu's ChatGLM series, and more. In the old 5,000-line catch-all, this was nearly impossible to do efficiently.

4. Open Source vs. Closed: An Invisible Frontline

China's aggressive open-source camp:

• DeepSeek — fully open across the board, from the 671B full R1 to 1.5B distilled mini models, with technical reports on arXiv and weights on Hugging Face
• Alibaba — a very aggressive open-source strategy; Qwen3.5-Plus is open with 397B total parameters and 17B active parameters, outperforming trillion-parameter closed models
• ByteDance — Seed-OSS-36B-Base is open, with a 512K context window and 12T training tokens
• Zhipu AI — continuous ChatGLM open-source releases

America's closed conservative camp:

• OpenAI — essentially closed since GPT-2; GPT-4, GPT-4o, o1, o3 are all black boxes
• Anthropic — the entire Claude line is closed; Opus 4.8 and Sonnet 4.5 are API-only
• Google — Gemini is closed; the occasional Gemma small models are exceptions, but flagship models never leave the nest
• xAI — Grok-1 was open-sourced once (314B MoE), but Grok 2 and Grok 3 immediately returned to closed

These are two different philosophies. The open route believes the model itself isn't the moat—application-layer innovation and ecosystem flourishing are; release your best model and the world applies it to healthcare, education, agriculture, and manufacturing. The closed route believes the model is the core competitive asset—hundreds of millions of dollars in compute and elite research teams are the barrier itself.

The data doesn't lie: Chinese vendors are visibly more aggressive and committed to open source than their American peers.

5. Context Windows: An Arms Race for Memory

If you understand AI competition through only one dimension, it should be the context window—how much a model can read in one go.

Early GPT-3 had 4K, roughly a short essay. GPT-4 reached 8K and 32K. GPT-4 Turbo jumped to 128K—about a 300-page novel. But the race didn't stop:

• Claude Opus 4.8: 1M tokens
• Gemini 2.0 Flash: 1M tokens
• Gemini 3.5 Flash: 1,048,576 tokens
• Qwen3.6-Plus: 1M tokens
• Qwen3.7-Max: 1M tokens
• Seed-OSS-36B-Base: 512K tokens

One million tokens means an AI can read a 700,000-character novel in one pass and remember every detail—who said what in which chapter, where a character's personality shifted, where the foreshadowing was planted.

Why does this matter? Because true intelligence isn't just about being clever—it's about remembering. Humans hold only 7±2 chunks in working memory; AI used to forget the beginning of a long paper by the time it reached the end. That limitation is breaking. When AI can process an entire book, an entire codebase, or a year of chat history at once, it can summarize personality changes from ten years of your diary, find a bug hidden for three years in a million-line project, or simultaneously handle legal contracts, case precedents, and client backgrounds like the best paralegal. The context-window arms race is fundamentally a contest for "long-term memory," a core of human cognition.

6. Distillation: A Master's Wisdom in a Student's Head

Another striking pattern in the data: the proliferation of distilled models.

The flagship DeepSeek-R1 has 671B parameters—a behemoth. But DeepSeek simultaneously released R1-Distill-Qwen models at 1.5B, 7B, 14B, 32B, and 70B.

Think of a Go grandmaster (the full model) who needs a huge room and table to show full strength. Distillation lets the master compress their skill into a young student (the small model). The student isn't as strong, but plays fast on a small board and far outclasses ordinary players. DeepSeek-R1-Distill-Qwen-1.5B is such a "student": only 1.5B parameters, runs on an ordinary laptop, yet inherits the 671B model's reasoning patterns in math, code, and logic.

The technique is knowledge distillation: the big model doesn't just tell the small model the answer—it teaches how it arrived at the answer, i.e., the distribution of the reasoning process. The small model learns the thinking, not just the result.

Distillation's significance is huge:

1. Democratization — not everyone has an A100 cluster; a 1.5B model runs smoothly on your MacBook 2. Lower cost — API pricing is per token; small models are far cheaper 3. Privacy — run locally, no data uploaded to the cloud

DeepSeek covers every deployment scenario: 1.5B for phones, 32B/70B for enterprise servers, 671B full-strength in the cloud.

7. Why Do This At All?

In 2023 there were roughly a dozen notable AI models; 2024 brought dozens; 2025 hundreds. In this dataset alone, openai.json runs nearly a thousand lines, alibaba.json over seven hundred, bytedance.json over five hundred.

Behind each model are millions of dollars in training costs, dozens of researchers, months or years of development. To users, they're just a pile of names—GPT-4o, Claude Opus, Gemini Pro, Qwen Max, DeepSeek-V3, Doubao-Seed-2.0—an ever-denser forest where you get lost the moment you step in.

The easy-learn-ai project blazes trails and plants signposts in that forest, answering the simplest questions of every developer and curious reader: Which models exist? Which country made them? Open or closed? What can they do? Where are the docs?

8. Final Words

The commit author, lishiqi.conard, appears to be from ByteDance (based on the email domain) and wrote plainly: "Ensure the accuracy and completeness of the data, making it easy for users to get the latest model information and usage guides."

But the value goes beyond "accurate and complete." At a time when AI models emerge daily, information curation is itself a creative act. The person who assembles fragments into a map matters as much as the one who discovers new continents—because no one can find direction in the fog until someone lights the lighthouse first.

A 5,000-line catch-all, split into nineteen clear files. That is lighthouse work.

---

*This article is based on commit e6c189a of the easy-learn-ai project, written in a Feynman-style explanation so readers without a technical background can follow the threads of the AI world.*