When the Model Catalog Outgrew the Phone Book: easy-learn-ai's Modular Refactor

Forum topic · 小凯 · 2026-08-23

Summary

The open-source AI encyclopedia project easy-learn-ai split a single 5,005-line JSON file containing all model data into 19 vendor-specific JSON files, one per company. The data loader now uses Webpack's dynamic require.context to auto-load every JSON file, so adding a new vendor requires zero code changes—just drop a new file in the directory. The refactor reflects the AI industry's scale in 2026: nine of the nineteen vendors are mainland Chinese companies (Alibaba, Baidu, ByteDance, DeepSeek, Kuaishou, MiniMax, Moonshot, Tencent, Zhipu AI), most large models now use MoE architectures (e.g., DeepSeek-V4-Pro with 1.6T total / 49B active parameters), and context windows have become an arms race, with Kimi K2.5 reaching 2.56M tokens. The commit was submitted by ByteDance engineer lishiqi on a Sunday night, illustrating that even cataloging model information now requires deliberate architecture—signaling the ecosystem's shift from startup era to industrial era.

Background

easy-learn-ai is an open-source project that helps ordinary people understand AI models—it doesn't train models or ship products; it just organizes and presents them. But a recent commit reveals something bigger about the state of the AI industry.

Originally, all model data lived in a single JSON file of 5,005 lines, alongside two smaller files for image-generation and video models. That worked when models were few. But by 2026, Alibaba alone ships nearly thirty models (Qwen3.5-Plus, Qwen3.7-Max, Qwen3-VL, QwQ-32B, Qwen-Image, Wan2.5...), DeepSeek iterates from R1 to V4, and Google, Anthropic, Baidu, ByteDance, Kuaishou, Meta, and Moonshot all maintain growing families. Editing, reviewing PRs, or adding a model to a 5,000-line JSON had become an organizational problem, not a technical one.

The Fix: Split by Vendor

The refactor splits the data into 19 JSON files, one per vendor, each containing models from the same "family."

The clever part is in the loader: instead of importing each file manually, it uses Webpack's dynamic module loading to grab every JSON file in the models directory and concatenate them into a list. The practical result:

> Adding a new vendor requires zero lines of code change—just drop a new JSON file into the directory. No imports, no registration, no loader edits. The folder is the API.

This is "convention over configuration," the kind of calm maturity mature projects earn.

What the Data Reveals About the Industry

Chinese vendors make up half the map

Nine of the nineteen vendors are from mainland China: Alibaba, Baidu, ByteDance, DeepSeek, Kuaishou, MiniMax, Moonshot, Tencent, and Zhipu AI. In 2023, a model encyclopedia would have been ~80% American companies; now China and the US are nearly even, with Chinese vendors leading in some areas (open-source video generation, long context).

The open/closed "split personality"

Alibaba ships both open (Qwen3.5-Plus, 397B MoE, Apache license) and closed (Qwen3.7-Max, 1M context) models. DeepSeek goes further—open-sourcing everything from 1.5B distills to 1.6T-total-parameter V4-Pro. The "open-source for traffic + closed-source for revenue" dual-track playbook is now standard among Chinese LLM companies.

MoE is the new baseline

Qwen3.5-Plus: 397B total / 17B active
DeepSeek-V4-Pro: 1.6T total / 49B active
DeepSeek-V4-Flash: 284B total / 13B active
Wan2.2: 27B total / 14B active

In 2026, a "large" model without Mixture of Experts is hard to take seriously.

The context-window arms race

| Model | Context window | |-------|----------------| | Qwen3.6-Plus | 1M tokens | | DeepSeek-V4-Pro | 1M tokens | | Kimi K2.5 | 2.56M tokens | | Claude 4 | 200K+ tokens |

1M tokens ≈ three million Chinese characters—enough to ingest the entire Three-Body trilogy, or a hundred contracts for risk-clause extraction.

Why This Small Commit Matters

The commit was pushed by ByteDance engineer lishiqi at 22:55 on a Sunday night—no flashy feature, just a thick phone book quietly split into nineteen thin volumes. But the signal is clear:

> AI models have multiplied to the point where cataloging them requires industry-grade infrastructure.

Good architecture is architecture that makes it easier for others to contribute: open a vendor's JSON to learn about a model, create a new file to add one, submit a PR without ever reading the whole project.