When Libraries Learn to Shelve Themselves: Restructuring the Model Universe of easy-learn-ai

Forum topic · 小凯 · 2026-07-18

Summary

This forum post from zhichai.net analyzes a major data-architecture refactor in the easy-learn-ai project (commit e6c189a). Previously, all AI model metadata lived in three capability-based files: model.json (5,005 lines for text models), img.json (667 lines), and video.json (491 lines). The refactor replaced these with about 20 company-centric files (openai.json, alibaba.json, anthropic.json, etc.), covering roughly 180+ models from Chinese and American labs. The author argues this shifts the cognitive framework from 'what a model can do' to 'who made it,' revealing company-level patterns: Alibaba's aggressive open-source Qwen matrix (including Qwen3.5-Plus, 397B MoE, priced at ~1/18 of Gemini 3 Pro), OpenAI's GPT empire reaching a 1,050K token context window, DeepSeek's reasoning-first philosophy (R1 trained with pure RL; DeepSeek-Math-V2 scoring 118/120 on Putnam), and ByteDance's full-modality Seed/Doubao suite. A standardized schema (modelName, company, openSourceStatus, contextWindow, modelTags, parent) replaced binary capability classification with a multi-tag system, better suited to today's all-in-one models. The post frames the refactor as turning model 'chaos' into a lineage-based map of the AI ecosystem.

When Libraries Learn to Shelve Themselves: Restructuring the Model Universe of easy-learn-ai

Key points

The refactor (commit e6c189a): The easy-learn-ai project replaced three capability-based data 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 roughly 20 company-centric files under src/data/models/.
New structure (one file per vendor): 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.
Core argument: capability-based classification (text/image/video) breaks down as models become multimodal and families grow; organizing by creator better preserves each company's technical "DNA," evolutionary lineages, and open-source vs. closed-source strategy.

Highlights from the new model files

OpenAI — the "GPT empire" (30 models, longest file)

GPT mainline: GPT-4 → GPT-4.1 → GPT-5 → GPT-5.1 → GPT-5.2 → GPT-5.5
Codex branch for agentic coding; o-series (o1/o3/o4-mini) for deep reasoning; DALL·E and GPT Image for images; Sora 1.0 → Sora 2 for video
GPT-5.5 offers a 1,050K token context window (~7 million Chinese characters); GPT-5.3-Codex scores 56.8% on SWE-Bench Pro
Almost entirely closed-source; its only open release, GPT OSS (120B/20B), came in August 2025

Alibaba — the "open-source matrix" (26 models)

Qwen family spans edge (Qwen3-4B-Instruct), mid-tier MoE (Qwen3-30B-A3B, 3B active), flagship (Qwen3.5-Plus: 397B total / 17B active, MMLU-Pro 87.8%), multimodal (Qwen3-VL, Qwen3-Omni), coding (Qwen3-Coder-480B), and generation (Qwen-Image, Z-Image, Wan2.2)
Qwen3.5-Plus API pricing is reportedly 1/18 of Gemini 3 Pro

DeepSeek — the "reasoning philosophy" (14 models)

DeepSeek-R1 approached OpenAI o1-1217 performance using pure reinforcement learning without human-annotated SFT data
DeepSeek-V4-Pro: 1.6T total parameters (49B active), dual Thinking/Non-Thinking modes
DeepSeek-Math-V2 reached gold-medal level at IMO 2025 and scored 118/120 on Putnam
DeepSeek-OCR reframes OCR as "contextual optical compression" with a 10:1 compression ratio

ByteDance — "full-modality loop" (18 models)

Doubao Seed 2.0 (pro/lite/code), vision models, Seedream 4.0 (4K images, up to 15-image composites), Seedance 2.0 (15-second multi-shot audio-video)
doubao-seed-1.6 supports thinking / non-thinking / auto modes with text, image, and video input; Seedance 2.0 accepts up to 9 images + 3 videos + 3 audio clips + natural language

Naming cultures

Chinese vendors (Qwen3.5-Plus, ERNIE-4.5-Turbo-128K, Doubao-Seed-1.6-Flash, DeepSeek-V3.2-Speciale) use dense numeric encoding — parameters, context length, and positioning packed into the name, building trust via specs.
US vendors (GPT-5, Claude Opus, Gemini Pro, Grok, Llama) favor personified, brand-style names.
Open-source vs. closed-source is a systematic strategic divide: Alibaba marks nearly all models open; Anthropic and OpenAI mark nearly all closed.

Technical observation: the schema metaphor

Each model entry now standardizes fields: modelName, company, country, openSourceStatus, releaseDate, description, modelTags, contextWindow, maxGenerationTokenLength, relatedLinks, and parent (series/parent model).

Crucially, there is no top-level "text vs. image vs. video" field — the multi-tag modelTags system (text generation, visual understanding, code enhancement, deep thinking, tool calling) composable combinations replace binary classification. A single model can be text-generating, vision-capable, code-enhanced, and tool-calling at once (e.g., Claude Opus 4.8, Gemini 3.5 Flash, doubao-seed-2.0-pro).

Data appendix (post-refactor scale)

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

Total: ~180+ models reorganized into 20 vendor files, covering major Chinese and American AI labs and startups.

Conclusion

Framed with the library metaphor: the old taxonomy shelved books by content ("this is a math book"); the new one shelves them by publisher ("this is from Commercial Press"). As models become hybrids of many capabilities, creator lineage provides the lasting cognitive anchor. The AI landscape of the future is not thousands of isolated tools but a few dozen model families — each with its own temperament, specialties, and followers. This refactor simply draws that ecosystem map a little more clearly.

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