When the AI Family Tree Gets Rewritten: A Classification Revolution in the Model Library
*Translation and summary of a zhichai.net forum post analyzing commit e6c189a of the easy-learn-ai project (July 12, 2026).*
Key points
- The refactor: The easy-learn-ai project — an effort to compile a comprehensive AI model reference — previously stored nearly 6,000 lines of JSON in just three files (
model.json,img.json,video.json), organized by capability. Developerlishiqi.conardsplit them into 19 vendor-based files covering 19 AI companies. - DeepSeek: ~20 models in
deepseek.json, including distilled variants (Distill-Qwen-1.5B/7B/14B/32B, Distill-Llama-8B/70B) whoseparentfields all point toDeepSeek-R1— a genealogical model family, mostly open source, with context windows from 64K to 128K. - Anthropic: Claude iterations (Opus 4.6, Opus 4.8, Sonnet 4.6, Haiku 4.5) emphasize agent coding, long-context stability, tool-trigger reliability, and honesty. Claude Opus 4.8 supports a 1M-token context window.
- Alibaba's dual strategy: Qwen3.5-Plus is open source (397B total / 17B activated parameters, MMLU-Pro 87.8%, ~1/18 the price of Gemini 3 Pro), while Qwen3.7-Max and Qwen3.6-Plus are closed-source, exclusive to Alibaba Cloud's Bailian/Model Studio with 1M-token context.
- Google: Gemini 2.0 Flash targets high throughput and low cost with 1M context and native multimodality; Gemini 3.1 Pro scores 77.1% on ARC-AGI-2; Gemini 3.5 Flash reportedly beats 3.1 Pro on several coding and agent benchmarks despite being a "Flash" tier.
- GPT-4 (2023): 8K → GPT-4o (2024): 128K → Gemini 2.0 Flash (2025): 1M → Claude Opus 4.8 (2026): 1M → Qwen3.7-Max (2026): 1M.
- The author argues that beyond ~1M tokens, AI shifts from "chat tool" to "research assistant."
- Open camp: DeepSeek (nearly all), Alibaba (Qwen3.5), Meta (Llama), parts of Zhipu AI, Stability AI.
- Closed camp: OpenAI, Anthropic, Google, Baidu, Moonshot, ByteDance (Seed).
- Some companies straddle both: Alibaba runs open Qwen3.5 alongside closed Qwen3.7.
Largest vendor files after the split
| File | Lines | Representative model families |
|---|---|---|
| openai.json | 981 | GPT series, o-series |
| alibaba.json | 752 | Qwen series |
| zhipu-ai.json | 570 | ChatGLM series |
| bytedance.json | 518 | Seed series |
| deepseek.json | 487 | DeepSeek-R1/V3 series |
| google.json | 421 | Gemini series |
| anthropic.json | 373 | Claude series |
| moonshot.json | 365 | Kimi series |
| baidu.json | 362 | ERNIE series |
| tencent.json | 353 | Hunyuan series |
Notable observations from the data
The context window arms race
Open vs. closed source
Capability tags as an evolution tracker
Model modelTags evolve from basic text generation, to visual understanding, to tool calling, with deep thinking/reasoning the hottest tag of 2025–2026 (DeepSeek-R1, Claude Opus, o-series).
The deeper meaning
The move from capability-based to vendor-based classification reflects an industry shift: a model's identity is moving from "what it can do" to "where it comes from." Choosing a model increasingly means choosing an ecosystem — OpenAI's API stack, self-hosting open DeepSeek, or Alibaba Cloud. Models are no longer isolated tools but entry points into expanding empires, and these 19 JSON files are the most honest snapshot of that map.
---
*Based on easy-learn-ai commit e6c189a.*