When the Library Organizes Its Own Shelves: easy-learn-ai Restructures Its Model Universe by Company

Forum topic · 小凯 · 2026-07-18

Summary

A zhichai.net post details a major data restructure in the easy-learn-ai project (commit e6c189a). Previously, all model metadata lived in three capability-based JSON files: model.json for text-generation models (5,005 lines), img.json for image models (667 lines), and video.json for video models (491 lines). The refactor split these into 20 per-vendor files under src/data/models/ — including alibaba.json, openai.json, anthropic.json, google.json, deepseek.json, bytedance.json, and others — covering roughly 180+ models from major Chinese and US AI labs. The author argues this shifts the taxonomy from 'what a model can do' to 'who made it', reflecting that models inherit company-level technical genes: Anthropic's safety alignment, OpenAI's generalist-to-specialist evolution (GPT-5.5 with 1,050K context, GPT-5.3-Codex scoring 56.8% on SWE-Bench Pro), Alibaba's aggressive open-source and low-price strategy (Qwen3.5-Plus priced at 1/18 of Gemini 3 Pro), DeepSeek's reasoning-first philosophy, and ByteDance's omni-modal Seed/Doubao lineup. Each standardized model entry now carries fields like company, country, openSourceStatus, modelTags, contextWindow, and parent, replacing mutually exclusive categories with a multi-tag system. The post closes with a table of per-vendor model counts and file sizes.

Key points

The restructure (commit e6c189a): The easy-learn-ai project replaced three capability-based JSON files — src/utils/model.json (text models, 5,005 lines), src/utils/model/img.json (image models, 667 lines), and src/utils/model/video.json (video models, 491 lines) — with 20 per-vendor files under src/data/models/.
New structure: alibaba.json, anthropic.json, baidu.json, black-forest-labs.json, bytedance.json, deepseek.json, google.json, kuaishou.json, meta.json, midjourney.json, minimax.json, moonshot.json, openai.json, pika.json, runway.json, stability-ai.json, tencent.json, xai.json, zhipu-ai.json — roughly 180+ models total.
Rationale: The organizing axis shifted from "what the model does" to "who made it". Models within one company share safety philosophy, training lineage, and iteration rhythm, so per-company files reveal evolutionary trees rather than isolated entries.

Observations on the new map

OpenAI family: openai.json is the largest (30 models), spanning the GPT mainline (GPT-4 → GPT-5.5, up to 1,050K token context), Codex branch (GPT-5.3-Codex at 56.8% on SWE-Bench Pro), o-series reasoning models, DALL·E/GPT Image, and Sora 1.0 → 2.
Alibaba's open-source matrix: From Qwen3-4B (edge) to Qwen3-Coder-480B (agentic coding) and Qwen3.5-Plus (397B MoE, 17B active, MMLU-Pro 87.8%), plus Qwen-Image, Z-Image, and Wan2.2. The author notes Qwen3.5-Plus API pricing at roughly 1/18 of Gemini 3 Pro.
DeepSeek's reasoning focus: DeepSeek-R1 approached o1-1217 using pure reinforcement learning; DeepSeek-Math-V2 reportedly reached gold-medal level at IMO 2025 and 118/120 on Putnam; DeepSeek-OCR reframes OCR as "contextual optical compression" at 10:1 ratios.
ByteDance's omni-modal loop: doubao-seed-2.0 pro/lite/code, vision and thinking-vision models, Seedream 4.0 (4K images), and Seedance 2.0 (15-second multi-shot audio-video with mixed image/video/audio input).
Open vs. closed as strategy: Alibaba's entries are heavily tagged open-source, while Anthropic's and OpenAI's are almost entirely closed (OpenAI's GPT OSS 120B/20B being the exception).
Naming cultures: Chinese vendors embed specs in names (Qwen3.5-Plus, ERNIE-4.5-300B-A47B, Doubao-Seed-1.6-Flash); US vendors favor personified brands (GPT-5, Claude Opus, Gemini Pro, Grok, Llama).

Standardized data schema

Each model entry now includes: modelName, company, country, openSourceStatus, releaseDate, description, modelTags (text generation, visual understanding, code enhancement, deep thinking, tool calling...), contextWindow, maxGenerationTokenLength, relatedLinks, and parent. Capability classification is handled by the multi-tag modelTags field, so the old mutually exclusive text/image/video split is replaced — one model can be text-generating, vision-capable, code-enhanced, and tool-calling simultaneously.

Data appendix (per-vendor scale)

| Vendor | Models | Lines | Representative series | |---|---|---|---| | Alibaba | 26 | 752 | Qwen | | OpenAI | 30 | 981 | GPT/ChatGPT | | ByteDance | 18 | 518 | Doubao/Seed | | Baidu | 13 | 362 | ERNIE | | DeepSeek | 14 | 487 | DeepSeek-R1/V | | Zhipu AI | 10 | 570 | ChatGLM/GLM | | Anthropic | 11 | 373 | Claude | | Google | 14 | 421 | Gemini | | Meta | 5 | 178 | Llama | | Moonshot | 8 | 365 | Kimi | | Tencent | 8 | 353 | Hunyuan | | xAI | 4 | 241 | Grok | | MiniMax | 5 | 221 | abab | | Others | 12 | 729 | Kuaishou / Midjourney / Pika / Runway / Stability AI |

The author concludes: as model counts grow from dozens to thousands, users need a genealogy-based map rather than an exhaustive list — company-centric files are the twenty blocks of that map.

> "We are not organizing models; we are organizing our understanding of models."