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

Reorganizing the AI Model Family Archive: easy-learn-ai Splits Monolithic model.json into Per-Vendor Files

Forum topic · 小凯 · 2026-08-13

Summary

The open-source easy-learn-ai project restructured its AI model catalog (commit e6c189a), splitting a single 5,000+ line model.json into 20 vendor-specific JSON files covering OpenAI, Anthropic, Google, Meta, xAI, Midjourney, Runway, Pika, Stability AI, Alibaba, Baidu, ByteDance, DeepSeek, Moonshot AI, MiniMax, Tencent, Kuaishou, Zhipu AI, and Black Forest Labs. Each model entry follows a standardized schema with fields such as modelName, company, country, openSourceStatus, releaseDate, contextWindow, maxGenerationTokenLength, and relatedLinks. Highlights include DeepSeek's MoE architecture (671B total/37B active parameters, V4-Pro at 1.6T total/49B active with a 1M-token context), Alibaba's Qwen3.5-Plus (397B total/17B active, up to 19x faster inference), GPT-5.5's 1.05M-token context window, and Gemini 3.5 Flash outperforming Gemini 3.1 Pro on coding benchmarks. The reorganization speeds up lookups, simplifies maintenance, and enables model comparison tools, recommendation systems, and search engines built on the structured data.

Reorganizing the AI World's "Household Registry" — The easy-learn-ai Model Family Archive Restructuring

How would you organize a registry of every AI model in the world? By release date? By country? By capability? By parent company? When facing dozens of vendors, hundreds of models, and text/image/video all in one family, the answer isn't obvious.

A Chaotic "Grand Encyclopedia"

In the easy-learn-ai open-source project, all model information used to live in one giant JSON file — over 5,000 lines, like an encyclopedia with no table of contents. The file, model.json, contained OpenAI's GPT series, Anthropic's Claude series, Google's Gemini series, China's DeepSeek, Alibaba's Qwen, Baidu's ERNIE, ByteDance, plus image generators like Midjourney, DALL·E, and Stable Diffusion, and video generators like Sora and Veo.

Imagine all residents of your apartment complex written on one big sheet of paper — no building number, no unit, no door number. Finding anyone means reading from top to bottom.

That was the state before this refactor.

The Art of Organizing: Archiving by "Family"

This update (commit e6c189a) did something simple but far-reaching: it split the encyclopedia into 20 small booklets, one per vendor.

US vendors — OpenAI, Anthropic, Google, Meta, xAI, Midjourney, Runway, Pika, Stability AI; Chinese vendors — Alibaba, Baidu, ByteDance, DeepSeek, Moonshot AI, MiniMax, Tencent, Kuaishou, Zhipu AI; and Germany's Black Forest Labs — each got their own dedicated archive.

Why does this matter?

1. Faster lookup. Want to know DeepSeek's models? Open deepseek.json. 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 links.

2. Easier comparison. Comparing US and China camps? Open alibaba.json and openai.json side by side; Qwen3.7-Max vs GPT-5.5 parameters, capabilities, and positioning are clear.

3. Simpler maintenance. Alibaba ships a new model? Update only alibaba.json without risking other vendors' data.

What's Inside the Archives

DeepSeek: The Open-Source Dark Horse

deepseek.json records 18 models, from DeepSeek LLM (January 2024) to DeepSeek-V4-Pro (April 2026). The highlight is the MoE (Mixture of Experts) architecture — 671B total parameters with only 37B activated per query, like a panel of experts where only relevant specialists answer each question. V4-Pro scales to 1.6T total / 49B active parameters with a 1,000,000-token context window — enough to read an entire long novel at once and remember every detail.

Alibaba Qwen: The All-Rounder's Toolbox

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

Qwen3.5-Plus stands out: 397B total parameters, only 17B active, up to 19x faster inference than its predecessor, and API pricing at 1/18 of Gemini 3 Pro.

OpenAI: The Closed-Source Kingdom's Precision Instruments

openai.json covers GPT-4 through GPT-5.5, o1 through o4-mini, DALL·E through Sora. Notable details: GPT-5.5 supports up to 1,050,000-token context windows, and GPT-5.3-Codex scored 56.8% on SWE-Bench Pro — meaning it can independently solve more than half of software engineering problems in that benchmark.

Google Gemini: The Multimodal Chameleon

google.json contains the Gemini family, possibly the most balanced group: from 2.0 Flash to 3.5 Flash, Nano Banana (image editing) to Imagen 4 (text-to-image), Veo 3 (video) to Gemini 3 Pro (all-round reasoning). Gemini 3.5 Flash beats its own Gemini 3.1 Pro on Terminal-Bench 2.1, MCP Atlas, and other coding/agent benchmarks while keeping Flash-series speed.

A Quiet "Standardization Movement"

Every model entry follows a unified schema:

  • modelName: model name
  • company: vendor
  • country: country
  • openSourceStatus: open or closed source
  • releaseDate: release date
  • description: capability description
  • modelTags: capability tags (text generation, vision, coding, tool calling, etc.)
  • contextWindow: context window size
  • maxGenerationTokenLength: max output length
  • relatedLinks: related links
This standardization means any developer or application can parse the data and build model recommendation systems, price-comparison tools, or capability search engines. Imagine a site where you type "I need a model good at Chinese, handling long documents, and cheap" and it instantly filters out Qwen-Flash, DeepSeek-V4-Flash, and ERNIE-4.5-Turbo with pros and cons.

For Everyone Else

Even if you're not a developer, this matters. AI models iterate monthly — even weekly. GPT-5.5 leads today; tomorrow Qwen3.7-Max may surpass it on some task; the day after, DeepSeek releases a new version at half the price. In this fast pace, information itself is competitiveness. Whoever knows fastest and most accurately which model best fits their needs gets the best results at the lowest cost.

Outlook

This refactor is just the beginning. Future additions could include pricing, inference speed, energy consumption, multilingual scores, safety ratings — even live API integration for one-click real-world benchmarking.

As the AI world grows more complex, "navigation maps" like easy-learn-ai become more important. After all, tools exist to serve people. No matter how powerful a model is, if you can't find it, use it, or choose it, it might as well not exist.

---

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

Tags

#easy-learn-ai#ai-models#open-source#json-data-structure#model-comparison#mixture-of-experts#deepseek#qwen

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/178633423