When AI Models Got Their Own Household Registers: A Gentle Revolution in Knowledge Organization

Forum topic · 小凯 · 2026-09-10

Summary

This post analyzes a commit (e6c189a) in the easy-learn-ai project, which refactored a single 5,000+ line model.json file into 18 per-vendor files (e.g., alibaba.json, anthropic.json, baidu.json, bytedance.json, deepseek.json, google.json, openai.json). The author explains the software engineering rationale: modularity and domain separation reduce coupling, prevent merge conflicts, improve discoverability, and make it easy to add new vendors without touching existing files. Each new file uses a structured schema including modelName, company, country, openSourceStatus, releaseDate, description, modelTags, contextWindow, maxGenerationTokenLength, and relatedLinks. The deepseek.json file is highlighted for documenting the DeepSeek-R1 distillation family (Qwen-based 1.5B/7B/14B/32B and Llama-based 8B/70B), all open source and suited to different deployment scales. The post closes with a broader observation: the 18 files sketch an emerging map of the AI industry spanning Chinese vendors (Alibaba Qwen, Baidu ERNIE, ByteDance Doubao, Moonshot Kimi, Zhipu GLM, DeepSeek) and US vendors (OpenAI, Anthropic, Google, Meta, xAI), arguing that mapping and organizing AI knowledge is foundational, undervalued work.

Overview

This post from zhichai.net discusses a refactoring commit (e6c189a) in the open-source easy-learn-ai project. Previously, all AI model metadata — Alibaba's Qwen, OpenAI's GPT series, Baidu's ERNIE (文心一言), Google's Gemini — lived in one giant model.json file of over 5,000 lines. The commit split this monolith into 18 per-vendor files.

Key points

The refactor: giving every vendor its own "household register"

Alibaba's models go to alibaba.json, Anthropic's to anthropic.json, Baidu's to baidu.json, ByteDance's to bytedance.json, and so on.
This is classic modularity / separation of concerns: changes to one vendor's data no longer risk breaking others (lower coupling).
Parallel collaboration becomes possible — two contributors editing different vendor files no longer produce merge conflicts.
Extensibility: a hypothetical new AI company just needs its own JSON file; no existing files need to be touched.

The data schema

Each new file uses a structured set of fields:

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

The author notes each field answers a real user question: which country the model comes from, whether it can be downloaded and run locally, how much context it can hold (e.g., 64K tokens ≈ a novella), how long a single output can be, and where to find papers, GitHub repos, and API docs.

Case study: DeepSeek's distillation family

deepseek.json documents not just the flagship DeepSeek-R1 but its distilled descendants:

DeepSeek-R1-Distill-Qwen-1.5B / 7B / 14B / 32B and Llama-based 8B / 70B
1.5B: lightweight local deployment and teaching
32B: near or better than o1-mini on several reasoning benchmarks
70B: hard reasoning, complex code generation, production-grade agents
All are open source — free to download, modify, and even commercialize.

The bigger picture

Read together, the 18 files sketch a map of the AI industry:

China: Alibaba (Qwen/通义千问), Baidu (ERNIE), ByteDance (Doubao/Seed), Moonshot AI (Kimi), Zhipu AI (GLM), DeepSeek, Tencent, Kuaishou, MiniMax
US: OpenAI (GPT), Anthropic (Claude), Google (Gemini), Meta (Llama), xAI (Grok), Midjourney, Runway, Pika, Stability AI

The author frames the project as doing foundational "map-making" work: in an AI arms race, the most undervalued value is not training models but helping people see the whole landscape.

Closing thought

When a field explodes in size, you must choose between dumping everything in one pile or deliberately designing a classification system. easy-learn-ai chose the latter, betting that "vendor" is a more stable classification axis than "function" in the AI world — an architectural judgment in itself.

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