Rebinding AI's Encyclopedia: How easy-learn-ai Restructured Its Model Database by Company

Forum topic · 小凯 · 2026-07-20

Summary

The easy-learn-ai project (commit e6c189a) refactored its AI model database from a single 5,000+ line model.json plus image and video JSON files into 19 per-company JSON files under src/data/models/, covering OpenAI, DeepSeek, Alibaba, Google, Anthropic, Meta, xAI, Moonshot, Zhipu AI, ByteDance, Tencent, Baidu, and others. Each model record follows a unified schema with fields for open-source status, release date, context window, capability tags, related links, and parent model relationships. Beyond the engineering change, the restructured data doubles as a snapshot of the 2025-2026 global AI landscape: OpenAI's march from GPT-4 to GPT-5.5 with 1M+ context and the surprise open-source GPT OSS release, DeepSeek's fully open-source lineage culminating in V4-Pro, Alibaba's sprawling Qwen family across text, vision, code, image, and video, Google's tiered Gemini strategy, and a stark split where Chinese vendors largely open-source while US vendors stay closed. This article explains the new architecture and what the data reveals.

> Source commit: e6c189a | easy-learn-ai project daily update

The easy-learn-ai project recently did something that looks mundane but is quietly significant: it split all model information out of one giant file into 19 independent, company-organized files — like turning a thick "global model phone book" into a set of company-by-company yearbooks.

From One Big Blob to Organized Ledgers

Before the refactor:

src/utils/model.json — a 5,000+ line monolith containing all text models
src/utils/model/img.json — image generation models (667 lines)
src/utils/model/video.json — video generation models (491 lines)

Every time OpenAI shipped a model, you had to dig through 5,000 lines of JSON; every version bump risked accidentally editing a neighboring company's data.

Now each company has its own cabinet under src/data/models/: 19 files covering alibaba, anthropic, baidu, black-forest-labs, bytedance, deepseek, google, kuaishou, meta, midjourney, minimax, moonshot, openai, pika, runway, stability-ai, tencent, xai, and zhipu-ai.

What the Data Reveals About the AI Landscape

OpenAI: from GPT-4 to GPT-5.5

Context window race: 8K (GPT-4) → 1M (GPT-4.1) → 1,050K (GPT-5.5)
A separate reasoning line: o1 → o3 → o4-mini; o3-deep-research can browse the web and write reports
August 2025: surprise open-source release of gpt-oss-120b and gpt-oss-20b (MoE, Apache 2.0)
Codex line (GPT-5 Codex → GPT-5.3-Codex) targets agentic coding; SWE-Bench Pro 56.8%

DeepSeek: open source + extreme engineering

Every entry in deepseek.json is open source. From DeepSeek LLM (Jan 2024) to the viral R1 (Jan 2025) to V4-Pro (Apr 2026, 1.6T total / 49B active parameters, 1M context). The R1 distill family (1.5B–70B on Qwen and Llama) lets individuals run a reasoning model locally. V3.2-Exp's DeepSeek Sparse Attention (DSA) cut API prices by more than half.

Alibaba: the Qwen universe

Text (Qwen3, 4B–235B, Plus/Flash/Max tiers), vision (Qwen3-VL), omni-modal (Qwen3-Omni-Flash, 119 text languages / 20 voice languages), code (Qwen3-Coder-480B MoE), image (Qwen-Image, Z-Image 6B/8-step), video (open-source Wan2.2, Wan2.5-Preview 1080P).

Google: steady, tiered Gemini

Flash = fast + cheap, Pro = strong + premium, Flash-Lite = value. Context windows start at 1M. Notably, Gemini 3.5 Flash reportedly beats Gemini 3.1 Pro on Terminal-Bench 2.1, MCP Atlas, GDPval-AA, and CharXiv Reasoning.

Anthropic: the cautious scholar

Claude's three pillars: hybrid reasoning (the model decides how long to think), long context (200K standard, 1M beta), and an emphasis on honesty. SWE-bench Verified 77.2% (Sonnet 4.5) and a ~144 Elo GDPval-AA lead over GPT-5.2.

Design Wisdom Behind the Architecture

1. A unified data contract: every model has the same fields — modelName/company/country, openSourceStatus, releaseDate, description, modelTags, contextWindow/maxGenerationTokenLength, relatedLinks, and parent (family lineage, e.g., DeepSeek-R1-Distill-Qwen-7B → DeepSeek-R1). Homogeneous data means the frontend renders everything uniformly. 2. Semantic capability tags across seven domains: text generation, deep thinking/reasoning, visual understanding, code enhancement, tool calling, image generation, and video generation. A model can carry multiple tags — GPT-5.5 covers the first five; Z-Image is a specialist. 3. Explicit family relationships: the parent field makes lineage instantly visible, like a biological taxonomy for models.

An Observation: The Open vs. Closed Divide

Reading all 19 files, one pattern is striking: Chinese companies overwhelmingly open-source (DeepSeek entirely, Alibaba's Qwen mostly, Zhipu GLM, Baidu ERNIE partially), while US companies overwhelmingly stay closed (Google, Anthropic, xAI; OpenAI except for GPT OSS). The line is commercial, not technical: Chinese vendors use open source for ecosystem positioning; US vendors treat API revenue as the moat. But OpenAI's August 2025 GPT OSS release suggests that line is softening — pressure from DeepSeek-style "open source + aggressive pricing" is forcing responses from the closed camp.

Epilogue: A Map That Grows Itself

After the refactor, adding a company means adding a JSON file; updating one vendor never touches the other 18; filtering by modelTags is trivial. Together, the 19 files form a panoramic snapshot of the 2026 AI industry — who builds general models, who targets verticals, who bets on open source, who stacks closed-source performance, who goes multimodal, and who goes deep on code.

Data architecture upgrades are never just about paying down technical debt — they are a re-modeling of domain knowledge.

---

*This article is based on analysis of data changes in easy-learn-ai commit e6c189a.*