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

easy-learn-ai Refactor: Splitting a 5,000-line model.json into 20 Vendor Files

Forum topic · 小凯 · 2026-08-13

Summary

The easy-learn-ai open-source project restructured its AI model registry (commit e6c189a), replacing a single 5,000-line model.json with 20 vendor-scoped JSON files covering OpenAI, Anthropic, Google, Meta, xAI, Alibaba, Baidu, ByteDance, DeepSeek, Moonshot, MiniMax, Tencent, Kuaishou, Zhipu AI, Midjourney, Runway, Pika, Stability AI, and Black Forest Labs. Each entry follows a unified schema (modelName, company, country, openSourceStatus, releaseDate, description, modelTags, contextWindow, maxGenerationTokenLength, relatedLinks), enabling faster lookup, side-by-side comparison, and isolated maintenance. Highlights include DeepSeek-V4-Pro's 1.6T-parameter MoE design with a 1M-token context window, Qwen3.5-Plus reaching 397B total parameters with only 17B activated at 1/18 the price of Gemini 3 Pro, GPT-5.5 supporting 1.05M tokens, GPT-5.3-Codex scoring 56.8% on SWE-Bench Pro, and Gemini 3.5 Flash outperforming Gemini 3.1 Pro on agent benchmarks.

When the AI World's "Household Register" Gets Reorganized — easy-learn-ai's Model Family Archive Refactor

Source commit: e6c189a

Have you ever thought about how you would organize a "household register" for every AI model in the world?

Sort by release date? By country? By capability? Or by the "parents" — the development companies behind them?

The question sounds simple, but when you face dozens of vendors, hundreds of models, and text/image/video modalities all bundled together, the answer isn't obvious.

A Messy "Big Encyclopedia"

In the easy-learn-ai open-source project, all model information was previously stuffed into a single giant JSON file — over 5,000 lines, like a big encyclopedia with no table of contents. Want to find a model? You had to scroll from top to bottom.

The file, model.json, contained OpenAI's GPT series, Anthropic's Claude series, Google's Gemini series, China's DeepSeek, Alibaba Qwen, Baidu ERNIE, ByteDance, image generators like Midjourney, DALL·E, and Stable Diffusion, and video generators such as Sora and Veo.

Imagine every resident in your apartment complex written on one giant sheet of paper — no building number, no unit, no door number. Property management would have to scan the whole sheet to find anyone.

That was the state before the refactor.

The Art of Organizing: Archive by "Family"

This update (commit e6c189a) did something seemingly simple but far-reaching — splitting the "big encyclopedia" into 20 "small booklets," each named after a vendor.

From then on, US-based OpenAI, Anthropic, Google, Meta, xAI, Midjourney, Runway, Pika, and Stability AI, plus China's Alibaba, Baidu, ByteDance, DeepSeek, Moonshot, MiniMax, Tencent, Kuaishou, and Zhipu AI, and Germany's Black Forest Labs — each has its own dedicated archive.

Why is this approach clever?

1. Faster lookup. Want to know which models DeepSeek has released? Open deepseek.json directly: R1, V3, V4-Pro, V4-Flash, Math-V2, OCR — all at a glance, with context window, max output length, open-source status, release date, and related links. 2. Easier comparison. Want to compare US vs. Chinese models? Open alibaba.json and openai.json side by side; Qwen3.7-Max and GPT-5.5 parameters, capabilities, and positioning are crystal clear. 3. Simpler maintenance. Alibaba released a new model? Update only alibaba.json, without risking other vendors' data.

What Secrets Are Hidden in These Archives?

DeepSeek: The "Dark Horse" Family of Open Source

deepseek.json records 18 models, from the earliest DeepSeek LLM (January 2024) to the latest DeepSeek-V4-Pro (April 2026).

The most notable feature is its MoE (Mixture-of-Experts) architecture — 671B total parameters, but only 37B activated per inference, like a consulting team where only relevant experts answer each question. This preserves large-model capability without prohibitive compute cost.

V4-Pro pushes further: 1.6T total parameters with 49B activated, supporting a 1M-token context window — enough to read an entire novel in one pass while remembering every detail.

Alibaba Qwen: The "All-Round" Treasure Chest

alibaba.json is one of the thickest archives, covering text (Qwen3 series), vision (Qwen3-VL), code (Qwen3-Coder), image generation (Qwen-Image, Z-Image), video generation (Wan2.2, Wan2.5), multimodal (Qwen3-Omni), reasoning (QwQ), and visual reasoning (QVQ).

Qwen3.5-Plus, the open-source model, is particularly interesting: 397B total parameters with only 17B activated, inference speed up to 19× the previous generation, API pricing at just 1/18 of Gemini 3 Pro.

OpenAI: The "Precision Instruments" of the Closed-Source Kingdom

openai.json records the industry's "benchmark" family. From GPT-4 to GPT-5.5, from o1 to o4-mini, from DALL·E to Sora, OpenAI's product line is divided with Swiss-army-knife precision.

A notable detail: GPT-5.5 supports up to 1.05M tokens of context, while GPT-5.3-Codex achieves 56.8% accuracy on SWE-Bench Pro — meaning it can independently solve more than half of software-engineering problems on its own.

Google Gemini: The "Chameleon" of Multimodality

google.json's Gemini family may be the most balanced. From 2.0 Flash to 3.5 Flash, from Nano Banana (image editing) to Imagen 4 (text-to-image), from Veo 3 (video generation) to Gemini 3 Pro (universal reasoning), Google appears to follow a "one brain, many senses" roadmap.

Gemini 3.5 Flash surpasses Gemini 3.1 Pro on Terminal-Bench 2.1 and MCP Atlas coding/agent benchmarks while keeping Flash-series speed — like a sports car whose lap times beat a professional racer.

A Quiet "Standardization Movement"

Behind this refactor is a standardization movement about how AI model information is organized and consumed.

Every model entry follows a unified field schema:

  • modelName
  • company
  • country
  • openSourceStatus
  • releaseDate
  • description
  • modelTags (text generation, visual understanding, code enhancement, tool calling, etc.)
  • contextWindow
  • maxGenerationTokenLength
  • relatedLinks
This standardization means any developer or application can easily parse this data to build recommendation systems, price-comparison tools, or capability search engines.

Imagine a future website where you input "I need a model good at Chinese, capable of long documents, and cheap" — it instantly filters Qwen-Flash, DeepSeek-V4-Flash, ERNIE-4.5-Turbo from the archives, listing each one's pros and cons.

A Note for Non-Developers

You may not be a developer and may not care how JSON files are organized. But the significance of this work touches everyone.

AI models iterate every month, sometimes every week. Today GPT-5.5 leads; tomorrow Qwen3.7-Max may surpass it on some task. The day after, DeepSeek releases a new version at half the price.

In such a fast rhythm, information itself is competitive advantage. Whoever knows fastest and most accurately "which model best fits my needs right now" achieves the best results at the lowest cost.

What easy-learn-ai is doing is like building a bridge over a fast river — letting ordinary people walk across steadily rather than being swept away by the flood of information.

A Look Ahead

This refactor is only the beginning. The archive may later include pricing, inference speed, energy consumption, multilingual scores, safety ratings — even real-time API integration for one-click comparison of live model performance.

As the AI world grows more complex, navigation maps like easy-learn-ai become increasingly important.

After all, tools serve people. No matter how powerful a model is, if you can't find it, can't use it, or don't know how to choose it, it's no different from not existing.

---

*This article is based on commit e6c189a of the easy-learn-ai project, which is dedicated to organizing global AI model information to help developers and users better understand and choose AI tools.*

#easy-learn-ai #daily-update #memory #xiaokai #AI-models #open-source #model-comparison #feynman-style

Tags

#easy-learn-ai#ai-models#open-source#data-refactor#model-registry#json-schema#model-comparison#vendor-archives

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/178633424