English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

easy-learn-ai Splits a 5,005-Line Model Catalog Into 19 Vendor-Based Files

Forum topic · 小凯 · 2026-08-23

Summary

easy-learn-ai, an open-source project cataloging AI models for the public, refactored its monolithic 5,005-line JSON data file into 19 vendor-specific files. The refactor used Webpack's require.context-style dynamic loading so that adding a new model vendor requires only dropping a new JSON file into the directory—zero code changes to the loader. The 19 files reveal industry trends: 9 of 19 vendors are from mainland China (Alibaba, Baidu, ByteDance, DeepSeek, Kuaishou, MiniMax, Moonshot, Tencent, Zhipu AI); dual open-source/closed-source strategies are standard among Chinese labs; Mixture-of-Experts (MoE) architectures are now the norm (e.g., Qwen3.5-Plus with 397B total/17B activated parameters, DeepSeek-V4-Pro with 1.6T total/49B activated); and context windows have become an arms race, with Kimi K2.5 reaching 2.56M tokens and several models offering 1M tokens. The commit, submitted by a ByteDance engineer on a Sunday night, signals that AI model ecosystems have grown large enough to require infrastructure-grade data architecture.

Overview

A daily-update commit in the easy-learn-ai project—a website that helps ordinary people understand AI models—did something deceptively mundane: it split a single 5,005-line JSON file containing all model data into 19 separate files. The post argues this refactor is a mirror reflecting the state of the AI industry in 2026.

The Problem: An Unusable "Model Yellow Pages"

Originally, all model metadata lived in one JSON file (5,005 lines), plus separate files for image-generation and video models. With the explosion of AI models—Alibaba alone has nearly thirty (Qwen3.5-Plus, Qwen3.7-Max, Qwen3-VL, QwQ-32B, Qwen-Image, Wan2.5, etc.), plus DeepSeek, Google Gemini, Anthropic Claude, Baidu ERNIE, ByteDance Seed, Kuaishou Kolors, Meta Llama, Moonshot Kimi—editing a single file meant scrolling for minutes, risking broken data, and reviewing multi-thousand-line diffs. As the author puts it: this is no longer a technical problem, but an organizational one.

The Fix: Split by Vendor + Dynamic Loading

The refactor split the data into 19 files, one per company/organization. The data loader no longer imports files manually; it uses a Webpack feature (require.context-style) to automatically scan the directory and concatenate all JSON files into one list:

> "Grab every .json file in that folder, parse them into an array, and merge into one big list."

Consequence: adding a new vendor requires zero code changes—just drop a new JSON file in the directory. This is a "convention over configuration" philosophy: the folder is the API.

What the Data Reveals About the Industry

China holds half the map

Nine of the 19 vendors are mainland Chinese: Alibaba, Baidu, ByteDance, DeepSeek, Kuaishou, MiniMax, Moonshot, Tencent, and Zhipu AI. In 2023, such a catalog would have been ~80% American companies; now China and the US are roughly even, with China leading in areas like open-source video generation and long context.

The open/closed "split personality"

Alibaba ships both open (Qwen3.5-Plus, 397B MoE under Apache; Qwen-Image, 20B MMDiT) and closed models (Qwen3.7-Max with 1M context; Qwen Image Plus via commercial API). DeepSeek open-sources everything from 1.5B distilled models to the 1.6T-parameter V4-Pro. "Open-source for traffic + closed-source for revenue" is now the standard playbook for Chinese LLM companies.

MoE is the new baseline

  • Qwen3.5-Plus: 397B total / 17B activated
  • DeepSeek-V4-Pro: 1.6T total / 49B activated
  • DeepSeek-V4-Flash: 284B total / 13B activated
  • Wan2.2: 27B total / 14B activated at inference
Mixture-of-Experts delivers similar capability at roughly 1/10 the compute.

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 ≈ a 3-million-character Chinese novel—enough to ingest the entire *Three-Body Problem* trilogy.

Why It Matters

easy-learn-ai doesn't train models or ship products; it only organizes and presents. When its catalog outgrew 3 files and required dynamic modular loading, it signaled that AI models now number enough to demand infrastructure-grade management—an indicator of the ecosystem moving from its "pioneer era" to an "industrial era." Good architecture, the author notes, makes it easier for others to participate: contribute a new model by adding one conventionally-structured file and opening a PR.

Epilogue

The commit was submitted by lishiqi, an engineer at ByteDance, at 22:55 on a Sunday night. No flashy features—just quietly dividing an unflippable yellow pages into nineteen slim volumes. The author suggests that looking back at 2026's open-source AI history, this may be the year even "organizing model information" required purpose-built architecture.

Tags

#easy-learn-ai#open-source#ai-models#refactoring#webpack#mixture-of-experts#chinese-ai#context-window

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633900