When One AI Model Encyclopedia Splits into Twenty Family Trees

> Source commit: e6c189a — model information restructuring in the easy-learn-ai project

1. One File, Five Thousand Lines, No Longer Sustainable

Before the refactor, the easy-learn-ai project stored all AI model information in a single model.json file over 5,000 lines long — every GPT, Claude, Qwen, DeepSeek, and Midjourney entry crammed into one JSON array. Two sibling files made it worse: img.json (600+ lines for image generation) and video.json (~500 lines for video generation), totaling over 6,000 lines.

The fix was not simply splitting by chapters, but splitting by family tree — one file per vendor, organized under src/data/models/. Five thousand lines became twenty tidy files.

2. Twenty Families: A Map of Global Competition

US camp:

• openai.json (981 lines) — the largest family, home of GPT
• anthropic.json (373 lines) — Claude Opus, Sonnet, Haiku
• google.json (421 lines) — Gemini
• meta.json (178 lines) — open-source Llama
• xai.json (241 lines) — Grok
• midjourney.json (29 lines), runway.json (52 lines), pika.json (48 lines) — image/video specialists

Chinese camp:

• alibaba.json (752 lines) — Qwen, second only to OpenAI in size
• bytedance.json (518 lines) — Seed series
• deepseek.json (487 lines) — R1
• zhipu-ai.json (570 lines) — GLM
• moonshot.json (365 lines) — Kimi
• baidu.json (362 lines) — ERNIE
• tencent.json (353 lines) — Hunyuan
• minimax.json (221 lines), kuaishou.json (51 lines) — MiniMax and Kling

Europe and others:

• black-forest-labs.json (56 lines) — FLUX, from the original Stable Diffusion team
• stability-ai.json (137 lines)

3. Why OpenAI Is Still the Largest File

openai.json at 981 lines is roughly 1.3x Alibaba's 752. Not because OpenAI has the most models, but because it has the longest history and the most complex product lines: GPT, o-series reasoning models, DALL·E, Sora, Whisper, and Embeddings — generations that all remain in service.

DeepSeek's 487 lines tell a different story: the R1 family spans 1.5B to 671B parameters like a precision Russian nesting doll — small versions run on phones, the largest needs a data center. "One idea, many body sizes."

4. The Context Window Arms Race

The contextWindow field shows explosive growth:

• Early GPT-3.5 Turbo: 16K
• GPT-4: 8K/32K
• Claude 3: 200K
• Claude Opus 4.8 and Qwen3.7-Max / Qwen3.6-Plus: 1M tokens
• Kimi: long-context pioneer starting at 128K

Long context transforms AI from a "sentence-by-sentence translator" into an "analyst who reads the whole document." Chinese vendors are notably aggressive here — a natural fit for long-form Chinese content like web novels, legal contracts, and classical texts.

5. Open vs. Closed: The Wall Is Crumbling

A clear divide exists in the openSourceStatus field:

• Closed: OpenAI, Anthropic, Google Gemini, Baidu ERNIE, ByteDance Seed, Moonshot Kimi
• Open: Meta Llama, Alibaba Qwen, DeepSeek R1, Zhipu GLM, Black Forest Labs FLUX, Stability AI

But the binary is blurring: Alibaba open-sources strong-but-not-strongest versions, Google open-sources Gemma while keeping Gemini closed, DeepSeek open-sources R1 while charging for API access. DeepSeek-R1's early-2025 impact — near-frontier performance at a fraction of the training cost — showed that open models are closing the gap with closed ones.

6. From Text to Image to Video: Modal Divergence and Fusion

The refactor folded all three modalities into per-vendor files: DALL·E and Sora live next to GPT in openai.json; Midjourney, Runway, Pika, and Kuaishou's Kling each hold their own. The trend points toward omnni-models — GPT-4o's "o" stands for "omni," handling text, audio, and images together. Future AI won't split "text models" from "image models," just as humans don't split "eyes" from "mouth."

7. Why Split by Vendor, Not by Capability?

The developers chose vendor-based organization, recognizing that model competition is really ecosystem competition. Switching from GPT-4 to Claude means switching API styles, pricing, toolchains, and safety policies — an entire workflow, not just a model. Users choose a family, and each family has its own DNA:

• OpenAI: first-mover advantage, fullest ecosystem, increasingly closed
• Anthropic: safety-first positioning
• Google: infrastructure powerhouse, long context and native multimodality
• DeepSeek: extreme cost-performance, aggressive open source
• Alibaba: the most open-source-friendly Chinese vendor, complete Qwen ecosystem
• ByteDance: fast follower pushing multimodal Seed models

8. German Engineering and Chinese Speed

Black Forest Labs — only 56 lines, yet one of the most important open-source forces in image generation. Its FLUX.2 series uses Rectified Flow Transformer techniques; the team came from the original Stable Diffusion crew.

DeepSeek — 487 lines of disruption. R1 trains reasoning via reinforcement learning rather than supervised fine-tuning; its distilled series spans 1.5B to 671B; its API pricing is a fraction of competitors'.

Together they sketch two global temperaments: European engineering rigor and Chinese speed.

9. The Metaphor Behind the Refactor

Surface reason: smaller files, faster loading, easier maintenance. Deeper reason: the AI model world has gone from one pond to an ocean. Two years ago, discussing AI models meant discussing OpenAI. Today there are twenty vendors, each with multiple sizes, versions, and modalities. Knowledge organization must evolve accordingly.

This is a Cambrian explosion: simple life forms suddenly diversifying into shell-armored, swimming, burrowing species — coexisting, competing, and depending on each other.

10. What Should Users Care About?

For everyday users, the refactor reflects a real shift: more choices, more complexity. Decisions now involve modality, context length, open vs. closed, price, language strengths, and capabilities. The twenty JSON files serve as a living map of the AI world, marking each vendor's position, strength, and specialty.

Epilogue: A Map Still Growing

The map won't stay still — a nvidia.json or apple.json may appear next month, or some unknown startup may release a shocking new model. But this refactor established a coordinate system: every new model now has a file, a family, and a label to call home.

From chaos to order, from one tangled file to twenty clear family trees — that is the story of this commit.

---

*Based on commit e6c189a of the easy-learn-ai project, which builds a structured AI model knowledge base.*