How an Open-Source AI Project Restructured Its 5,000-Line model.json Into 20 Vendor Files

Forum topic · 小凯 · 2026-08-15

Summary

The commit e6c189a of the easy-learn-ai project replaces a single ~5,000-line model.json (plus img.json and video.json) with a vendor-centric directory of 20 JSON files under src/data/models/, covering 6,195 lines of data across Alibaba (Qwen), Anthropic (Claude), Baidu (ERNIE), DeepSeek, Google (Gemini), Meta (Llama), OpenAI (GPT), xAI (Grok), Zhipu (GLM), and others. Each model now carries a standardized record with modelName, company, country, openSourceStatus, releaseDate, contextWindow, maxGenerationTokenLength, modelTags, relatedLinks, and a parent field that traces distillation lineage. modelApi.ts uses webpackContext with regExp /\.json$/ to auto-discover every JSON in the directory, so adding a new vendor only requires dropping in a new file — no code changes. The article also maps the global AI landscape reflected in the data: 10 Chinese vendors, 7 U.S. vendors, and 3 European/other labs, and reflects on how convention-over-configuration, structured metadata, and open-source ideology together make the catalog more readable, maintainable, and extensible.

Key points

Refactor scope (commit e6c189a, 2026-07-12): the easy-learn-ai project replaced a monolithic model.json (~5,000 lines) along with separate img.json and video.json files with a vendor-centric directory src/data/models/ containing 20 JSON files totaling 6,195 lines.
Vendors covered: Alibaba, Anthropic, Baidu, Black Forest Labs, ByteDance, DeepSeek, Google, Kuaishou (Kling), Meta, Midjourney, MiniMax, Moonshot (Kimi), OpenAI, Pika Labs, Runway, Stability AI, Tencent, xAI, and Zhipu AI.
Standardized model record introduces fields such as modelName, company, country, openSourceStatus, releaseDate, contextWindow, maxGenerationTokenLength, modelTags, relatedLinks, and a parent pointer that captures distillation lineage (e.g. DeepSeek-R1-Distill-Qwen-1.5B → parent DeepSeek-R1).
Auto-loading via webpack: modelApi.ts uses import.meta.webpackContext("../data/models", { recursive: false, regExp: /\.json$/ }) to auto-discover and merge every vendor JSON at build time. Adding a new vendor requires only creating a new JSON file — no code edits.
Convention over configuration: the directory convention and file naming are sufficient for the loader; repetitive boilerplate is delegated to tooling so contributors can focus on data quality.
Snapshot of the global AI landscape revealed by the dataset:
China (10): Alibaba (Qwen), Baidu (ERNIE), ByteDance (Seed), DeepSeek, Kuaishou (Kling), MiniMax, Moonshot (Kimi), Tencent (Hunyuan), Zhipu AI (GLM) — plus Huawei (Pangu) and iFlyTek (Spark) referenced as missing.
United States (7): OpenAI (GPT), Anthropic (Claude), Google (Gemini), Meta (Llama), xAI (Grok), Midjourney, Stability AI.
Europe / others (3): Black Forest Labs (FLUX, Germany), Pika Labs, Runway.
Context-window examples cited: DeepSeek-R1 at 64K tokens; Claude Opus 4.8 at 1000K tokens (roughly the length of *Dream of the Red Chamber* plus the first half of *Journey to the West*).
Open vs. closed source pattern: Meta's Llama, Alibaba's Qwen, DeepSeek's R1, and Zhipu's GLM are highlighted as the de facto open-source infrastructure of the ecosystem, paralleling Linux's role on the web.
Why it matters: as of 2026 the answer to "which AI model should I use?" is a dense table rather than a single name; structured vendor files make the catalog easier to read, maintain, and extend for end users comparing open-source Chinese LLMs against closed frontier models.

Reference data

Project: easy-learn-ai
Commit: e6c189a
Date: 2026-07-12
Files touched: 20 vendor JSON files + modelApi.ts
Total: 6,195 lines across 20 AI vendors

> Note: the source post references a future-dated commit (2026-07-12) for a project that, per its README, was bootstrapped in early 2022; this summary reproduces those details as stated by the author without further verification.