When AI Models Get a 'Household Registry': Restructuring Knowledge by Vendor

Forum topic · 小凯 · 2026-09-06

Summary

A forum post on zhichai.net reviews a data restructuring of the open-source easy-learn-ai project, which reorganized AI model metadata from capability-based JSON files (text, image, video) into 19 vendor-based family files such as anthropic.json, deepseek.json, and alibaba.json. The author argues vendor-based organization preserves product lineages, compares flagship model specs including 1,000K-token context windows (Claude Opus 4.8, DeepSeek-V4-Pro, Gemini 2.5 Pro, Qwen3.7-Max) and output limits up to 384K tokens, and highlights design fields like contextWindow, maxGenerationTokenLength, modelTags, and parent lineage. Industry trends drawn from the data include a US-China bipolar landscape, MoE architectures enabling efficient small-parameter activation, deep-reasoning modes becoming standard, and multimodality moving from demos to production. The post ends with a beginner's guide to studying AI model families and cites the project's GitHub repository.

Overview

This post from zhichai.net discusses a data restructuring in the open-source easy-learn-ai project (github.com/ConardLi/easy-learn-ai, commit e6c189a). Previously, model data lived in a few large capability-based JSON files: model.json for text models, img.json for image models, video.json for video models. The restructuring reorganized the catalog into 19 vendor-based "family" files.

Key points

Vendor-based files now include: alibaba.json (Qwen family, Wan video models), anthropic.json (Claude Opus/Sonnet/Haiku tiers), deepseek.json (R1/V3/V4 lineage), google.json (Gemini, Veo), openai.json, plus baidu.json, bytedance.json, tencent.json, and others.
Rationale: AI models have genealogies. Grouping Claude 3.5 Sonnet through Claude 4.6 Sonnet, or DeepSeek's evolution from V2 (236B MoE) through R1, V3 (671B MoE), V3.1, V3.2-Exp (DSA sparse attention, 50% cost reduction), V4-Pro (1.6T total params, 1M context) and V4-Flash (284B, lightweight), reveals each vendor's product strategy at a glance.

New data structure fields

contextWindow (in K tokens): flagship models now reach 1000K (1M tokens) — Claude Opus 4.8, DeepSeek-V4-Pro, Qwen3.7-Max, Gemini 2.5 Pro. Roughly 750,000 Chinese characters, or three copies of *Dream of the Red Chamber*.
maxGenerationTokenLength: DeepSeek-V4-Pro supports 384K output (a novella), Claude Opus 4.8 128K, Gemini 2.5 Pro 65K, while earlier models offered only 8K–32K.
modelTags: text generation, deep thinking, vision, code, tool calling, image/video generation. Trends: generalist models (Claude Opus 4.8, Gemini 3.5 Flash, Qwen3.5-Plus) now hold all tags; reasoning modes have become standard across vendors; multimodal models like Qwen3-Omni (text/image/audio/video/speech) are squeezing out pure-text models.
parent: lineage field — e.g., DeepSeek-R1-Distill-Qwen-32B is distilled from DeepSeek-R1.

Industry trends read from the data

US–China bipolar structure: US vendors (OpenAI, Anthropic, Google, Meta, xAI, Midjourney, Runway, Pika, Stability AI) push closed-source flagship performance; Chinese vendors (Alibaba, DeepSeek, Baidu, ByteDance, Tencent, Moonshot, MiniMax, Zhipu AI, Kuaishou) push open-source and cost efficiency.
"Small parameters, big intelligence": MoE architectures activate only a fraction of total parameters (DeepSeek-V4-Pro: 49B of 1.6T; Qwen3.5-Plus: 17B of 397B), enabling consumer-hardware deployment.
Reasoning as core competency: Following DeepSeek-R1's 2025 RL breakthrough, adaptive thinking, Deep Think, and Thinking modes are now industry standard.
Multimodal goes practical: Veo 3.1 (1080P with sound), Wan2.5 (audio-video sync, 10s 1080P), Runway Gen-4 (60s video) are entering real ad, film, and e-commerce workflows.

A beginner's guide

1. Pick one family to study deeply — DeepSeek (open-source, well-documented) or Claude (closed-source, safety-focused). 2. Watch context window and output length — a 1000K/128K model and a 128K/8K model are entirely different creatures. 3. Try models hands-on — DashScope (Qwen3-Max), Anthropic Console (Claude Sonnet), DeepSeek's site (V3.2). 4. Track genealogies quarterly — Claude went from 3.5 to 4.8 in 18 months; DeepSeek from V2 to V4 in one year.

---

*Source: easy-learn-ai project (https://github.com/ConardLi/easy-learn-ai), commit e6c189a.*