When Someone Decided to Map the Entire AI World: The easy-learn-ai Model Database Refactor

Forum topic · 小凯 · 2026-07-26

Summary

This post from the zhichai.net tech forum recounts how the open-source easy-learn-ai project (commit e6c189a, July 2026) split a monolithic 5,000+ line model.json file into 20 vendor-specific files under src/data/models/, cataloging 230+ AI models from providers including OpenAI, Anthropic, Google, Meta, xAI, Alibaba (Qwen), DeepSeek, Zhipu AI, Baidu, ByteDance, Tencent, Moonshot, MiniMax, Kuaishou, Black Forest Labs, Midjourney, Runway, Pika, and Stability AI. The author highlights four industry trends visible in the data: context windows expanding from 8K to 1M tokens in three years, multimodality becoming standard, reasoning capabilities (e.g., DeepSeek-R1 chain-of-thought) emerging as the new battleground, and the widening yet converging split between closed models (OpenAI, Anthropic, Google, Baidu, xAI) and open-source ones (DeepSeek, Qwen, Llama, GLM). The refactor makes the dataset sustainable to maintain and easier for newcomers and developers to search, compare, and cite model specs.

When Someone Decided to Map the Entire AI World

> Source: easy-learn-ai project, commit e6c189a > Date: 2026-07-12

Imagine living on a new continent that expands every day—new mountains rise, new rivers shift, new cities appear overnight. You hold a map, but it is a single parchment crammed with everything: towns, mountains, rivers, forests all squeezed together, colors overlapping, borders blurred. Want to find the city of "DeepSeek"? You have to read from beginning to end. Want to know the distance between "Claude" and "Gemini"? You have to decipher the chaos yourself.

That is not a fable. That has been the real situation of anyone trying to understand the world of AI models over the past few years.

The Parchment Problem

In the easy-learn-ai open-source project, all AI model information long lived in a single file: src/utils/model.json—over 5,000 lines. It contained OpenAI's GPT series, Anthropic's Claude family, Google's Gemini corps, China's DeepSeek, Qwen, ERNIE, and more, all dumped into one bucket like mixed building blocks.

Updating a single model meant searching 5,000 lines. Comparing two vendors meant constant scrolling and Ctrl+F. And the count kept climbing—past 100, past 200, toward 300. So someone decided it was time to draw a real map.

From Chaos to Order

Commit e6c189a did something seemingly simple but far-reaching: split the giant file into 20 independent files, one per vendor, under src/data/models/. Like a cartographer finally drawing one map per country instead of cramming all borders onto one sheet.

US camp: the closed-source giants

OpenAI — 38 models, the largest territory. From GPT-4 (8K context, 2023) to GPT-5.5 (1M context, 2026). GPT-4o's "omni" design handles text, audio, images, and video. Its tiered strategy (GPT-4.1 nano/mini/standard) assigns roles like a team: nano for lowest latency and cost, mini for balance, standard for maximum intelligence.
Anthropic — 12 models, a boutique approach. Claude naming maps to length: Opus (flagship, 1M context, 128K output), Sonnet (value pick), Haiku (fast, cheap, real-time). Claude 4.8 (May 2026) introduced "adaptive thinking"—answering simple questions quickly while slowing down for hard ones.
Google — 15 models, firmly betting on multimodality. Gemini 2.0 Flash supports 1M context across text, image, audio, video, and PDF; Gemini 3.1 Pro scores 77.1% on ARC-AGI-2; Gemini 3.5 Flash reportedly beats 3.1 Pro on coding, agentic, and multimodal benchmarks while staying fast.
Meta — 7 models. The Llama series may be the most influential open-source release of recent years, from Llama 2 to Llama 4.
xAI — 9 models. Grok features real-time X-platform data, 1M-token context, and a deliberately humorous personality.

China camp: dual-track open and closed

Alibaba (Qwen) — 28 models. Qwen3.5-Plus has 397B total parameters but only 17B active thanks to MoE—like a hospital with 397 departments where only 17 consult per patient. Its API price is reportedly 1/18 of Gemini 3 Pro.
DeepSeek — 17 models, open-source reasoning benchmark. DeepSeek-R1 (early 2025) shipped not only the main model but distilled versions from 1.5B to 70B—like a Go master sharing game records with students of every level, all free.
Zhipu AI — 24 models. The GLM series uses an autoregressive blank-filling architecture distinct from GPT's pure autoregression; the line spans mobile-runnable GLM-Z1-1B to flagship GLM-4.5-Plus with 1M context.
Baidu (ERNIE) — 13 models, knowledge-enhanced. ERNIE-5.0 is natively omni-modal, aimed at enterprise Q&A, coding, and content creation.
Second tier — ByteDance (Seed), Tencent (Hunyuan), Moonshot (Kimi, famous for long Chinese context up to 200K early on), MiniMax (abab), Kuaishou (Kling, strong in video generation).

Vertical specialists

Black Forest Labs (FLUX image models), Midjourney (art-scene icon via Discord), Runway (Gen-3 text-to-video), Pika (fast, easy video), Stability AI (Stable Diffusion, open-source image generation).

Four Trends in the Data

1. Context windows exploding

GPT-4 (2023): 8K. GPT-4o (2024): 128K. GPT-4.1 (2025): 1M. A 125x jump in three years—from a short story to the entire Harry Potter series. Models can now swallow thousands of pages of contracts or entire codebases in one pass.

2. Multimodality as standard

Flagships—GPT-4o, Claude Opus 4.8, Gemini 3.5 Flash, Qwen3.7-Max—all support vision; many add audio. Not a bonus feature but a fundamental shift toward how humans process the world.

3. Reasoning as the new battleground

Before 2024, competition was about knowledge; after 2025, about thinking. DeepSeek-R1, Claude Opus 4.8, ERNIE-X1, and the GPT-o series all reason step-by-step. R1's chain-of-thought brought it near OpenAI o1 on math and coding. The exam moved from open-book to closed-book.

4. Open vs. closed

The dataset includes an openSourceStatus field. Closed camp: OpenAI, Anthropic, Google, Baidu, ByteDance, xAI—black boxes. Open camp: DeepSeek, Alibaba/Qwen, Meta/Llama, Zhipu/GLM—weights and methods fully public. Closed vendors treat models as revenue; open vendors treat them as ecosystem foundations. And the gap is closing fast: R1 matches o1 on several benchmarks; Qwen3.5-Plus outperforms trillion-parameter closed models.

Why Draw This Map?

Each model entry contains a relatedLinks field with three links—official announcement, technical docs, GitHub repo. That is curation, not data dumping. In an age of information explosion, finding the right information matters more than owning it. The refactor makes this curation sustainable: new model? Add an entry to the vendor file. No more needle-in-haystack searches through 5,000 lines.

What the Timeline Shows

2023: GPT-4 arrives; the era of large models begins.
2024: Multimodality explodes (GPT-4o, Gemini 1.5).
2025: Reasoning models rise (DeepSeek-R1, Claude 3.5 Sonnet).
2026: Context windows pass one million; agents become the keyword—models start "doing," not just "talking."

Progress is not linear but leaping: each breakthrough—multimodality, long context, reasoning—opens entirely new applications.

Closing Thought

"The map is not the territory." No matter how precise, a map is an abstraction. The JSON files capture parameters, context lengths, release dates—but not the subjective feel of Claude's honesty, GPT's fluency, or DeepSeek's depth. Even so, a good map tells you where you are, what surrounds you, and what may lie ahead. On a continent that changes daily, having a map beats not having one.

---

*Based on data from easy-learn-ai commit e6c189a. The project catalogs 230+ AI models from 20 vendors across text generation, vision understanding, coding, and tool use.*