When AI Models Pile Up Like Supermarket Shelves — One Developer Built Them a World Map

Forum topic · 小凯 · 2026-08-31

Summary

A Chinese forum post discusses the easy-learn-ai open-source project, which recently restructured its AI model catalog from a single 5,000-line JSON file into 18 separate files, one per vendor — OpenAI, Google, Anthropic, Meta, xAI, DeepSeek, Alibaba, ByteDance, Baidu, Moonshot (Kimi), MiniMax, Zhipu AI, Tencent, Kuaishou, and others. The post uses this refactor as a lens to explain the 2023–2026 explosion in large language models: layered model tiers (flagship, value, lightweight, domain-specific), narrowing gaps between open-source models (Llama 4, DeepSeek-R1, Qwen3.5) and closed ones, a context-window arms race reaching 1M+ tokens (with Llama 4 Scout at 10M), and the collective rise of Chinese vendors in long context, open-source ecosystems, pricing, and multimodal capabilities. It argues that with hundreds of models and rapid release cycles, organized information — a reliable 'map of AI models' — matters as much as any individual model, helping users and developers choose the right tool for the right scenario.

When AI Models Pile Up Like Supermarket Shelves — Someone Built Them a "World Map"

> Source commit: e6c189a

The Problem: Information Explosion Without Organization

In 2023, knowing the major AI models fit on one sheet of paper: OpenAI's GPT, Google's Bard, and a few academic projects. By 2025–2026, hundreds of large language models exist worldwide — open and closed, each specialized for coding, long documents, images, or video — with version names stacking like Russian nesting dolls (Qwen3 → 3.5 → 3.7; Claude 4 → 4.5 → 4.8).

The easy-learn-ai open-source project, built to make AI understandable for ordinary people, originally stored its entire model catalog in one massive JSON file — over 5,000 lines. Model relationships (e.g., DeepSeek-R1 as a "parent" of many distilled variants) existed only in maintainers' heads, not in any structured form.

The Refactor: 5,000 Lines Split into 18 Files

A recent commit split the monolithic file into 18 independent files, one per vendor:

US: OpenAI, Google, Anthropic, Meta, xAI, Midjourney, Runway, Pika, Stability AI, Black Forest Labs
China: Alibaba (Qwen), Baidu (ERNIE), ByteDance (Doubao/Seed), DeepSeek, Moonshot (Kimi), MiniMax, Zhipu AI, Tencent, Kuaishou

Like reorganizing a bookstore's books from one giant box onto shelves by publisher.

Highlights from the Shelves

OpenAI: GPT-4 → Turbo → 4o → 4.1 → GPT-5 → 5.1. GPT-4o ("omni") pioneered true multimodality; GPT-4o mini targets cheap, high-volume workloads.
Anthropic: Claude's Opus (flagship, 1M-token context in Opus 4.8), Sonnet (best value, beat GPT-5.2 on coding benchmarks in Sonnet 4.6), Haiku (fastest, real-time).
Google: Gemini's huge context windows (1M+ tokens ≈ 750K Chinese characters, roughly two copies of *Dream of the Red Chamber*), plus unique integration with Search, Docs, and Gmail.
Meta: Llama's open "recipe" lets startups run near-GPT-4 models locally. Llama 4 Maverick uses a Mixture-of-Experts architecture (400B total, 17B active per token — like hospital specialists consulting only when needed); Llama 4 Scout supports a 10M-token context window.
DeepSeek: R1 matched OpenAI o1 in math/code reasoning at a fraction of the cost, then distilled its reasoning into small models — a 1.5B-parameter distill beat GPT-4o on math competitions. Fully open-source.
Alibaba: Qwen3.5-Plus (397B total / 17B active) outperforms trillion-parameter rivals at ~1/18 of Gemini 3 Pro's API price; Qwen3.7-Max handles 1M-token input and 64K output.
Moonshot: Kimi bet everything on long context — from 200K characters at launch to 2M, dominating legal-document analysis, knowledge management, and research use cases.
ByteDance: Doubao/Seed models are the most "engineering-driven," backed by data from Douyin/Toutiao; Seed-OSS-36B is open-source with 512K context.
Baidu: ERNIE integrates Baidu's knowledge graph into training, reducing factual hallucination; ERNIE-5.1-Thinking supports 128K input/output.

Four Trends Behind the Restructure

1. Models are layering: flagship (Claude Opus, GPT-5, Qwen3.7-Max) → value (Sonnet, Gemini Flash) → lightweight (Haiku, DeepSeek-Distill-1.5B) → domain-specific (Qwen-Coder, Doubao-Seed-Code). Choose models like phone plans — fit, not price. 2. Open vs. closed are converging: DeepSeek-R1, Llama 4, and Qwen3.5 now match or beat some closed models, pushing enterprises toward hybrid strategies (local open models for sensitive data, cloud closed models for general tasks). 3. Context-window arms race: from GPT-4's 8K tokens in 2023 to 1M+ tokens in 2026 — enabling whole-book comprehension, full-codebase analysis, and hours-long meeting transcription. 4. Chinese vendors rising collectively: nearly half the 18 files are Chinese companies, leading in long context, open-source downloads (Qwen), price competition (DeepSeek's API briefly at ~1/50 of OpenAI's), and multimodal generation.

What It Means for You

Ordinary users: more choices, lower prices — free tiers for most daily tasks, Kimi for long documents, Doubao for media generation, Claude for coding, Gemini for the Google ecosystem.
Developers/enterprises: build a model-selection capability — match models to scenarios across cost, security, latency, and capability, rather than chasing the newest release.

The developer who split 5,000 lines into 18 files didn't invent a breakthrough technology. But in an era of information explosion, he built order — and a reliable map matters more than any single model, because you can't choose well if you don't know what your options are.

---

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