easy-learn-ai Project Restructures Model Registry into 20 Vendor-Based JSON Files

Forum topic · 小凯 · 2026-08-13

Summary

The open-source easy-learn-ai project has split its monolithic model.json—which previously mixed data from dozens of AI vendors into a single 5,000+ line file—into 20 vendor-specific JSON files such as openai.json, deepseek.json, and alibaba.json. Each entry follows a unified schema capturing modelName, company, country, openSourceStatus, releaseDate, description, modelTags, contextWindow, maxGenerationTokenLength, and relatedLinks. This restructure preserves detailed records for major families including OpenAI's GPT-5.5, DeepSeek's V4-Pro with 1.6T parameters and 1M-token context, and Alibaba's Qwen3.5-Plus. The standardized format enables developers to build model comparison tools, recommendation engines, and pricing dashboards, making it easier for users to navigate the rapidly evolving AI landscape.

A community Chinese-language post discusses commit e6c189a of the easy-learn-ai open-source project, which reorganized its sprawling AI model registry into a cleaner, vendor-based structure.

Background

Before the change, every model's metadata lived in a single model.json file exceeding 5,000 lines. The file mixed OpenAI, Anthropic, Google, DeepSeek, Alibaba, Baidu, ByteDance, Midjourney, Stability AI, and many others into one undifferentiated document, making it hard to locate, compare, or update specific models.

Key Changes

The registry is split into roughly 20 vendor-named JSON files:
USA: OpenAI, Anthropic, Google, Meta, xAI, Midjourney, Runway, Pika, Stability AI
China: Alibaba, Baidu, ByteDance, DeepSeek, Moonshot (月之暗面), MiniMax, Tencent, Kuaishou, Zhipu AI
Germany: Black Forest Labs
Each file lists that vendor's models with fields such as modelName, company, country, openSourceStatus, releaseDate, description, modelTags, contextWindow, maxGenerationTokenLength, and relatedLinks.

Highlights From the Archives

DeepSeek (deepseek.json) — 18 models from DeepSeek LLM (Jan 2024) to DeepSeek-V4-Pro (Apr 2026). V4-Pro uses MoE with 1.6T total parameters, 49B active, and a 1M-token context window.
Alibaba (alibaba.json) — One of the largest files, covering Qwen3 text models, Qwen3-VL vision, Qwen3-Coder, Qwen-Image, Z-Image, Wan2.2/Wan2.5 video, Qwen3-Omni, QwQ reasoning, and QVQ visual reasoning. Qwen3.5-Plus is open-source with 397B total / 17B active parameters and reportedly up to 19× faster inference, priced at roughly 1/18 of Gemini 3 Pro.
OpenAI (openai.json) — Tracks GPT-4 through GPT-5.5, o1 through o4-mini, DALL·E, and Sora. GPT-5.5 supports up to 1.05M tokens; GPT-5.3-Codex reaches 56.8% on SWE-Bench Pro.
Google (google.json) — Gemini 2.0 Flash through 3.5 Flash, Imagen 4, Veo 3, and Gemini 3 Pro. Gemini 3.5 Flash reportedly exceeds Gemini 3.1 Pro on Terminal-Bench 2.1 and MCP Atlas while retaining Flash-series speed.

Why the Standardization Matters

With a uniform schema, any developer can parse the dataset to build comparison tools, recommendation systems, or pricing dashboards. The author suggests a future interface where users query "a Chinese-friendly, long-context, low-cost model" and receive ranked candidates such as Qwen-Flash, DeepSeek-V4-Flash, or ERNIE-4.5-Turbo.

Outlook

The post frames the change as a small but foundational step in making AI model information consumable, suggesting future additions like real-time pricing, inference speed, energy use, multilingual scores, and safety ratings.