When the AI World Explodes, Someone Quietly Made a Map

This post walks through a commit in the easy-learn-ai project (commit e6c189a) that reorganizes a bloated AI model catalog into a structured, queryable database.

The Problem: a 5,000-Line Dump

The project's src/utils/model.json had grown to over 5,000 lines, cramming every model — OpenAI's GPT series, DeepSeek's R1 family, Google's Gemini, Alibaba's Qwen, xAI's Grok, Stability AI's image models — into one unmanageable file. With models iterating weekly, maintaining it had become nearly impossible.

The Fix: 19 Per-Provider Files

The commit deletes the monolithic file and replaces it with a src/data/models/ directory containing 19 files:

• alibaba.json — Alibaba's Qwen family
• anthropic.json — Anthropic's Claude family
• baidu.json — Baidu's ERNIE series
• black-forest-labs.json — FLUX image generation
• bytedance.json — ByteDance's Seed and Doubao
• deepseek.json — the full DeepSeek lineup
• google.json — the Gemini lineage
• plus kuaishou.json, meta.json, midjourney.json, minimax.json, moonshot.json, openai.json, pika.json, runway.json, stability-ai.json, tencent.json, xai.json, zhipu-ai.json

A Standardized Schema

Each model now has a uniform identity card with fields:

• modelName — e.g. "DeepSeek-R1", "Claude Opus 4.8"
• company and country
• openSourceStatus — open vs. closed
• releaseDate, description
• modelTags — text generation, deep reasoning, vision, code, tool use
• contextWindow — 64K, 128K, up to 1M tokens
• maxGenerationTokenLength
• relatedLinks — docs, papers, GitHub repos
• parent — e.g. DeepSeek-R1-Distill-Qwen-7B → DeepSeek-R1

With this schema, queries like "list all open-source Chinese models" become a simple filter over the files.

Patterns in the Data

Open-source push (mostly Chinese vendors):

• DeepSeek — fully open, from 671B R1 down to 1.5B distills
• Alibaba — Qwen3.5-Plus: 397B total params, 17B active
• ByteDance — Seed-OSS-36B-Base with a 512K context window and 12T training tokens
• Zhipu AI — ongoing ChatGLM open releases

Closed-source approach (mostly US vendors):

• OpenAI — essentially closed since GPT-2
• Anthropic — the entire Claude family is API-only
• Google — Gemini closed; occasional Gemma small-model releases
• xAI — Grok-1 was open-sourced once; Grok 2/3 returned to closed

The Context-Window Arms Race

Context window — how much a model can read at once — has become a key competitive dimension:

• Claude Opus 4.8: 1M context
• Gemini 2.0 Flash: 1M context; Gemini 3.5 Flash: 1,048,576 tokens
• Qwen3.6-Plus and Qwen3.7-Max: 1M context
• Seed-OSS-36B-Base: 512K context

A 1M-token window means an AI can read an entire novel — or codebase — in one pass and recall every detail, transforming tasks like finding long-hidden bugs or summarizing years of records.

Knowledge Distillation

The data also shows a surge of distilled models, exemplified by DeepSeek-R1-Distill-Qwen at 1.5B, 7B, 14B, 32B, and 70B parameters. In knowledge distillation, the small model learns not answers but the reasoning style of the 671B teacher model. Benefits:

1. Democratization — a 1.5B model runs on a laptop 2. Lower cost — far cheaper inference than large-model APIs 3. Privacy — local deployment keeps data on-device

Why It Matters

By 2025–2026, notable AI models number in the hundreds — openai.json alone is nearly 1,000 lines, alibaba.json 700+, bytedance.json 500+. The project answers the basic questions every developer has: which models exist, who made them, whether they're open, what they can do, and where to find their docs.

The commit's author, lishiqi.conard (a ByteDance engineer, judging by the email domain), wrote plainly in the commit message: "ensure data accuracy and completeness, making it easy for users to get the latest model information and usage guides." But in a field where models emerge daily, structured information curation is itself a creative act — like lighting a lighthouse in the fog.

---

*Based on easy-learn-ai project commit e6c189a.*