One Encyclopedia Splits into Twenty Family Trees: What a Single Commit Reveals About the AI Landscape

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

*This is a structured English summary of the original Chinese forum post.*

Key points

• The easy-learn-ai project refactored a monolithic model.json (~5,000+ lines) plus img.json (~600 lines) and video.json (~500 lines) into twenty per-vendor JSON files under src/data/models/.
• The split was done by vendor (family tree), not by modality, reflecting that AI competition is ecosystem-level, not model-level.

The twenty files as a map of the industry

US camp:

• openai.json (981 lines) — largest; GPT, o-series, DALL·E, Sora, Whisper, Embeddings
• google.json (421), anthropic.json (373), meta.json (178), xai.json (241), midjourney.json (29), runway.json (52), pika.json (48)

China camp:

• alibaba.json (752) — Qwen, second only to OpenAI
• zhipu-ai.json (570), bytedance.json (518), deepseek.json (487), moonshot.json (365), baidu.json (362), tencent.json (353), minimax.json (221), kuaishou.json (51)

Europe/others:

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

Notable findings from the data

• Why OpenAI's file is biggest: not the most models, but the longest history and most complex product lines — GPT, o-series (reasoning), DALL·E (image), Sora (video), Whisper (speech), Embeddings — all coexisting across generations.
• DeepSeek's structured scaling: R1 and its distilled series span 1.5B to 671B parameters — one idea, many sizes, from phone-scale to datacenter-scale, trained with reinforcement learning rather than supervised fine-tuning, at a fraction of competitors' API pricing.
• Context-window arms race: GPT-3.5 Turbo at 16K → GPT-4 at 8K/32K → Claude 3 at 200K → Claude Opus at 1M → Qwen Max at 1M. Chinese vendors are especially aggressive on long context, driven by real demand for long-form Chinese text (novels, legal contracts, financial reports).
• The open/closed wall is cracking: pure-closed (OpenAI, Anthropic, Google Gemini, Baidu ERNIE, ByteDance Seed, Moonshot Kimi) vs. pure-open (Meta Llama, Qwen, DeepSeek-R1, GLM, FLUX, Stable Diffusion) — but hybrid strategies are emerging (Qwen Max closed while Qwen Plus is open; Google's Gemma open, Gemini closed).
• Modality convergence: text, image, and video models now live inside the same vendor files, mirroring the shift to native multimodality (GPT-4o's "omni", Gemini, Qwen).

Why split by vendor instead of by capability

Choosing a model means entering an ecosystem — API style, pricing, tooling, safety policy. The vendor-based structure acknowledges that users pick a "family," not a single model:

• OpenAI: first-mover advantage, broadest ecosystem, increasingly closed
• Anthropic: safety-first positioning
• Google: infrastructure scale, long context, native multimodality
• DeepSeek: extreme cost-efficiency, aggressive open-sourcing
• Alibaba: most open among Chinese vendors, most complete Qwen ecosystem
• ByteDance: fast follower pushing multimodal Seed models

Takeaway

The refactoring is more than an engineering optimization — it is an acknowledgment that the AI world has gone from "one pond" to "an ocean." There is no single "best model" anymore, only the model best suited to your scenario. The map will keep growing (a future nvidia.json or apple.json may be coming), but the commit establishes the coordinate system for cataloging it — an archive of the industry's Cambrian explosion.

---

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