When AI Learns to Sort: What a Database Refactor Reveals About the AI Model Landscape

The 5,000-Line Suitcase

Imagine moving into a new apartment and finding that the previous tenant stuffed every piece of clothing — winter coats, summer T-shirts, suits, pajamas, even swim trunks — into one giant suitcase. All you need is a pair of socks, but you have to turn the whole thing inside out.

That's what developers of easy-learn-ai — an open-source website helping ordinary people learn about AI models — faced with their core asset: a model database where every model's name, company, capability tags, context window, and documentation links were dumped into a single model.json file that was 5,005 lines long.

For a machine, 5,000 lines is trivial. For humans, it meant scrolling through thousands of lines to check DeepSeek's latest model, carefully avoiding edits to a neighboring company's data, and suffering nightmarish Git merge conflicts. Worse, image generation and video generation models were mixed in with text models — like going to the supermarket for milk and finding tires and screwdrivers on the shelf.

So the developers did something simple but wise: they tidied the room.

From One Big Box to Twenty Small Drawers

The refactor's core logic: split by vendor. The single model.json became 20 independent JSON files, one per AI company:

• alibaba.json — Alibaba (Qwen/Tongyi series)
• anthropic.json — Anthropic (Claude series)
• baidu.json — Baidu (ERNIE series)
• black-forest-labs.json — Black Forest Labs (FLUX image generation)
• bytedance.json — ByteDance (Seed series)
• deepseek.json — DeepSeek
• google.json — Google (Gemini series)
• kuaishou.json — Kuaishou
• meta.json — Meta (Llama series)
• midjourney.json — Midjourney
• minimax.json — MiniMax
• moonshot.json — Moonshot AI (Kimi)
• openai.json — OpenAI (GPT series)
• pika.json — Pika Labs (video generation)
• runway.json — Runway (video generation)
• stability-ai.json — Stability AI
• tencent.json — Tencent (Hunyuan series)
• xai.json — xAI (Grok series)
• zhipu-ai.json — Zhipu AI (ChatGLM series)

Image models (img.json) and video models (video.json) were also organized into their own directories: models/ for text, img/ for images, video/ for video — like a library's classification index. The engineering principle behind this is Separation of Concerns: don't mix different things together, just as you wouldn't keep socks and cookware in the same drawer.

What the Data Reveals: A Map of the AI World

After the cleanup, the 20 files contain 240 AI models — and the numbers tell the story of an ongoing silent war.

The Giant: OpenAI's Moat

OpenAI tops the list with 38 models, iterating from GPT-3.5 Turbo (2023) to GPT-5 (2026). Its strategy: cover every price point and every scenario — GPT-5 for peak performance, GPT-5 mini for value, GPT-4o mini and GPT-4.1 nano for lightweight use, GPT-4o for vision, GPT-4 Turbo for pinned versions. Like a restaurant serving everything from street food to Michelin three-star cuisine, OpenAI wants to leave no one an excuse to say "I don't need OpenAI."

Chinese Players: Nine Companies, 100+ Models

Nine of the 20 vendors are Chinese, together exceeding a third of all models:

• Alibaba (28 models): Qwen iterated from Qwen2 to Qwen3.7-Max, covering all price tiers like OpenAI — but more aggressive on open source. The open-sourced Qwen3.5-Plus is a 397-billion-parameter mixture-of-experts model claimed to outperform trillion-parameter closed models, with API pricing at 1/18 of Gemini 3 Pro.
• Zhipu AI (24 models): The ChatGLM series was among the earliest open-sourced Chinese LLMs, from ChatGLM-6B to GLM-4, proving Chinese teams can make models small and fast enough for ordinary developers.
• ByteDance (20 models): The low-profile Seed series powers AI features in apps like Douyin, with heavy investment in efficiency.
• DeepSeek (17 models): worth a closer look.

DeepSeek's "Distillation Magic"

Of DeepSeek's 17 models, 11 are distilled variants. Distillation: a top student (DeepSeek-R1) summarizes its problem-solving methods into teaching notes for apprentices of different aptitudes (from Qwen-1.5B's 1.5B parameters to Llama-70B's 70B). R1 used its 671-billion-parameter "superbrain" to generate 800,000 high-quality reasoning samples, then trained small models on them. The result: a 1.5B-parameter Qwen model distilled from R1 reportedly matches GPT-4 on math reasoning.

The significance: AI "intelligence" can be separated from "body size." You don't need everyone to own an elephant to carry heavy loads — the elephant draws the route map, and a donkey following it reaches the destination. DeepSeek open-sourced all 11 distilled models, meaning any developer with a gaming GPU can run a near-R1-level reasoning model locally — unthinkable a year ago.

Anthropic's Long-Context Obsession

Anthropic has fewer models (12), but they share one trait: ultra-long context windows. Claude Opus 4.8 supports 1M tokens — roughly 750,000 Chinese characters, about two copies of *Dream of the Red Chamber*. Anthropic's bet: future AI competition is not about who is smarter, but who "remembers" more. An AI that can read ten years of your company's financial reports, emails, and meeting notes gives advice based on *your specific situation*, not general knowledge.

Images and Video: The New Battlefield

• Midjourney: one model, but its position in AI art is like Coca-Cola's in cola — the brand *is* the category.
• Black Forest Labs (FLUX): two models, but FLUX.1 shook up open-source image generation, proving open models can rival Midjourney in many scenarios.
• Runway and Pika: the twin pillars of video generation (Runway Gen-4, Pika 2.0) — the leap from "generating pictures" to "generating films."
• Kuaishou: its Keling model generates coherent motion video; Chinese teams are not falling behind on this emerging track.

The Code in the Tags: What Can AI Actually Do?

Each model carries capability tags that sketch AI's ability boundaries:

• Foundation: text generation — the "default skill" of virtually all 240 models.
• Advanced: deep thinking — models like DeepSeek-R1 and Claude Opus "think" before answering via Chain-of-Thought, writing out reasoning steps to boost accuracy on complex problems.
• Vision: seeing and describing — GPT-4o, Claude, Gemini, and Qwen series can interpret menus, circuit diagrams, and photos.
• Tools: doing real work — the closest to the "Agent" concept. Tool-calling models execute operations (weather APIs, calculators, browsers). Claude Sonnet 4.6 goes furthest, clicking, typing, and browsing like a human.
• Code: programmer's assistant — a contested battleground; every vendor touts SWE-bench scores, and each 1% gain means fewer all-nighters for developers.

Open Source vs. Closed Source: Two Roads Collide

The openSourceStatus field reveals a clear divide:

• Closed camp: OpenAI, Anthropic, Google, Baidu (ERNIE), ByteDance (Seed), Moonshot (Kimi), MiniMax, Tencent (Hunyuan)
• Open camp: Meta (Llama), Alibaba (Qwen), DeepSeek, Zhipu AI (ChatGLM/GLM), Stability AI, Black Forest Labs (FLUX)

The pattern is intriguing: among US giants, OpenAI and Anthropic are firmly closed and Google semi-closed — but Meta goes the opposite way, fully open-sourcing Llama. China is split: Alibaba, Zhipu, and DeepSeek chose open source; Baidu, ByteDance, and Moonshot chose closed.

Why open source? Because in AI, the ecosystem matters more than the model itself. Meta's open Llama earned free optimization and adaptation from the whole community; Alibaba's open Qwen became one of the world's most-used open model families. The closed logic also holds: a model is core competitive capital — open-sourcing GPT-5 could vaporize half of OpenAI's $20 billion valuation overnight.

Neither strategy is absolutely right or wrong, but one thing is certain: open source democratizes AI capability. Without Llama and Qwen, ordinary developers might only ever call someone else's API, paying per token with no control over their own data.

Context Windows: AI's "Memory Length"

The easily overlooked but critical contextWindow field measures how much AI can "remember" at once:

• 64K: ~50,000 characters, a thin book
• 128K: ~100,000 characters, *The Little Prince*
• 256K: ~200,000 characters, a novella
• 1000K (1M): ~750,000 characters, two copies of *Dream of the Red Chamber*

Early GPT-3.5 had only 4K — it started "forgetting" after a dozen conversation turns. Today's Claude Opus 4.8 at 1M tokens can ingest a full year of work emails to write your annual summary. Expanding context windows is the key step upgrading AI from "chat tool" to "work partner."

Conclusion: Those Who Tidy the Room See the Whole Picture First

This refactor's significance goes beyond "cleaner code." By categorizing 240 models by vendor, the developers effectively drew a map of the AI world. On it we see:

• Chinese and American camps each holding half the territory
• OpenAI and Anthropic building high walls around closed models
• Meta, Alibaba, and DeepSeek breaking barriers with open source
• Image and video generation opening new battlefields
• A context-window arms race redefining capability boundaries
• Distillation freeing AI intelligence from parameter scale

A refactor meant only to improve maintainability became a window onto the entire AI industry: competition and collaboration, openness and closure, giant and miniature, memory and forgetting.

That's the charm of code: you tidy a room, and discover you've tidied the whole world.

---

Data source: easy-learn-ai project commit e6c189a Model count: 20 vendors, 240 models Vendors covered: OpenAI, Anthropic, Google, Meta, xAI, Stability AI, Midjourney, Runway, Pika Labs, Black Forest Labs, Alibaba, Baidu, ByteDance, DeepSeek, Zhipu AI, Moonshot AI, MiniMax, Tencent, Kuaishou