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

> Source commit: e6c189a

Key points

• The open-source easy-learn-ai project split a single 5,000+ line JSON file of AI model data into 18 vendor-specific files (6,000+ total lines), making the catalog maintainable and reflecting model lineage (e.g., DeepSeek-R1 and its distilled variants).
• The reorganization surfaces four industry trends: model tiering, open-source catching up with closed models, the context-window arms race, and the rise of Chinese AI vendors.

The problem: model overload

Back in 2023, knowing the major models took one sheet of paper: OpenAI's GPT, Google's Bard, and a few academic entries. By 2025–2026, there are hundreds of LLMs from different countries and companies — open and closed, code-focused, long-document readers, image and video generators — with version names stacking like Russian nesting dolls (Qwen3, Qwen3.5, Qwen3.7; Claude 4, 4.5, 4.8). Information exploded, but information organization didn't keep up.

The refactor: from one box to 18 shelves

The original 5,000-line JSON was like every book in a bookstore crammed into one cardboard box. The maintainer split it by "publisher" — one file per AI 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

OpenAI

From GPT-4 (2023) through GPT-4 Turbo → GPT-4o → GPT-4.1 → GPT-5 → GPT-5 mini → GPT-5.1. GPT-4o's "o" stands for "omni" — the first true multimodal model (text, image, audio). GPT-4o mini is the lean, cheap, fast option for high-volume tasks like support bots and classification.

Anthropic

Claude's tiers are named Opus / Sonnet / Haiku — flagship, price-performance, and fastest, respectively. Claude Opus 4.8 handles 1M-character context (enough to summarize the whole of *The Three-Body Problem*), and Claude Sonnet 4.6 reportedly beats GPT-5.2 on coding benchmarks.

Google

Gemini models feature enormous context windows: Gemini 2.0 Flash and 3.5 Flash both support 1M+ tokens (~750,000 Chinese characters). Gemini also integrates with Google Search, Docs, and Gmail — ecosystem integration OpenAI can't yet match.

Meta

The Llama series publishes full "recipes" (data, architecture, weights). Llama 4 Maverick has 400B total parameters but activates only 17B per token — a Mixture-of-Experts (MoE) design, like a hospital calling only the relevant specialist. Llama 4 Scout's context window reaches 10M tokens.

DeepSeek

DeepSeek-R1 matched OpenAI o1 on math and coding reasoning at a fraction of the training cost, and was fully open-sourced. Its "distillation" — teaching R1's reasoning method to small models (1.5B–70B, including Qwen and Llama bases) — let a 1.5B-parameter model beat GPT-4o on math competitions.

Alibaba

Qwen3.5-Plus: 397B total parameters, 17B activated, outperforming trillion-parameter rivals on benchmarks at ~1/18 the API price of Google Gemini 3 Pro. Qwen3.7-Max supports 1M-character input and 64K-character output. Qwen is among the most-downloaded open-source model families globally.

Moonshot (Kimi)

Doubly focused on long context from day one: 200K characters initially, now 2M — enough to answer pinpoint questions about the entire *Journey to the West*.

ByteDance

Seed-OSS-36B is open-source with 512K-character context; Doubao-Seed-1.6 is multimodal (text, image, video) with tool calling. ByteDance leverages massive content data from Douyin/Toutiao and embeds AI directly into its products.

Baidu

ERNIE integrates Baidu's knowledge graph into training, reducing factual hallucination. ERNIE-5.1-Thinking supports 128K input and 128K output with deep-reasoning capability, and is fused with Baidu Search.

Four trends behind the reorganization

1. Model tiering: flagship (Claude Opus, GPT-5, Qwen3.7-Max), value tier (Sonnet, Gemini Flash), lightweight (Haiku, GPT-4o mini, DeepSeek-Distill-1.5B), and domain-specific (Doubao-Seed-Code, Qwen-Coder). Choose models like phone plans — fit over price. 2. Open vs. closed converging: DeepSeek-R1, Llama 4, and Qwen3.5 now match or beat some closed models on benchmarks. Enterprises increasingly adopt hybrid strategies — local open-source for sensitive data, cloud closed APIs for general tasks. 3. Context-window arms race: from GPT-4's 8K tokens (2023) to 1M+ tokens (Gemini, Claude) — enabling whole-book QA, codebase analysis, hour-long meeting transcription review, and contract risk audits. 4. Chinese vendors rising collectively: nearly half of the 18 vendors are Chinese, leading in long context (Kimi 2M chars, ERNIE 128K output), open-source downloads (Qwen), cost (DeepSeek APIs once ~1/50 of OpenAI's), and multimodality (Doubao, Tongyi Wanxiang).

What it means for you

For everyday users, model abundance means more choice at lower cost — a free DeepSeek tier for routine tasks, Kimi for long documents, Claude for coding and deep research, Gemini for the Google ecosystem. For developers and enterprises, it demands real model-selection capability: weighing scenario, cost, safety, and latency rather than chasing the newest release.

Closing thought

Today's champion may be overtaken in three months. In such an environment, a reliable information curator matters more than any single powerful model — because you can't choose well if you don't know the options. Splitting 5,000 lines into 18 files isn't a technical breakthrough, but building order in an age of information explosion is itself a form of value.

---

*This article is based on commit e6c189a of the easy-learn-ai project.*