When AI Models Get a Family Registry: Restructuring the Knowledge Order of Model Data

Forum topic · 小凯 · 2026-09-06

Summary

The easy-learn-ai open-source project restructured its AI model catalog from capability-based JSON files (text, image, video) into 19 vendor-family files such as alibaba.json, anthropic.json, deepseek.json, google.json, and openai.json. This vendor-oriented organization preserves product lineages (e.g., Claude 3.5 to 4.6 Sonnet, DeepSeek V2 to V4) and makes competitive strategies legible. Key structured fields include contextWindow, with frontier models like Claude Opus 4.8, DeepSeek-V4-Pro, Qwen3.7-Max, and Gemini 2.5 Pro reaching 1000K tokens; maxGenerationTokenLength, where DeepSeek-V4-Pro leads with 384K output; modelTags, showing reasoning, vision, coding, tool use, and multimodal generation becoming standard; and parent, capturing distillation lineage. Aggregated data reveals a US-China bipolar landscape (OpenAI/Anthropic/Google vs. Alibaba/DeepSeek/Baidu/ByteDance), Mixture-of-Experts enabling sparse activation (DeepSeek-V4-Pro: 1.6T total, 49B active), reasoning modes becoming universal, and multimodal video generation (Veo 3.1, Wan2.5, Runway Gen-4) entering production workflows. Source: https://github.com/ConardLi/easy-learn-ai

Key points

The easy-learn-ai project (https://github.com/ConardLi/easy-learn-ai, commit e6c189a) refactored its AI model database from a few large capability-based files (model.json, img.json, video.json) into 19 vendor-family files — effectively giving every AI vendor a "household registry."

The new organization

alibaba.json — Qwen family (Qwen3-Max, Qwen3-Omni) plus Wan video models
anthropic.json — the Claude lineup from Opus to Haiku
deepseek.json — R1/V3/V4 series, a timeline of Chinese open-source AI
google.json — Gemini and Veo
openai.json, baidu.json, bytedance.json, tencent.json, and more

Vendor-based grouping preserves product genealogy — e.g., Claude 3.5 Sonnet → 4 → 4.5 → 4.6 Sonnet, or DeepSeek LLM → V2 (MoE pilot) → R1 → V3 (671B MoE) → V3.1 → V3.2-Exp (DSA sparse attention, 50% cost cut) → V4-Pro → V4-Flash.

Designed-in data fields

contextWindow: frontier models (Claude Opus 4.8, DeepSeek-V4-Pro, Qwen3.7-Max, Gemini 2.5 Pro) now reach 1000K tokens (~750,000 Chinese characters) — infrastructure for agentic workloads.
maxGenerationTokenLength: DeepSeek-V4-Pro leads with 384K output; Claude Opus 4.8 at 128K; Gemini 2.5 Pro at 65K; older models 8K–32K.
modelTags: text generation, deep thinking, vision, code enhancement, tool calling, image/video generation. "All-rounder" models (Claude Opus 4.8, Gemini 3.5 Flash, Qwen3.5-Plus) now collect every tag; reasoning modes have become table stakes; multimodality is swallowing text-only models.
parent: captures lineage, e.g., DeepSeek-R1-Distill-Qwen-32B descends from DeepSeek-R1.

Industry trends read from the data

US–China bipolar structure: US labs (OpenAI, Anthropic, Google, Meta, xAI, Midjourney, Runway, Pika, Stability AI) push closed 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 sparse activation — DeepSeek-V4-Pro activates 49B of 1.6T total; Qwen3.5-Plus activates 17B of 397B.
Reasoning as core competence: post-DeepSeek-R1, adaptive thinking / Deep Think / Thinking modes are now standard.
Multimodal goes practical: Veo 3.1 (1080P with sound), Wan2.5 (audio-visual sync, 10s 1080P), Runway Gen-4 (60s video) are entering ads, film, and e-commerce workflows.

Beginner's path suggested by the author

1. Pick one model family to study deeply (DeepSeek or Claude recommended). 2. Watch context window and output length — they define what a model can do. 3. Try models hands-on (DashScope, Anthropic Console, DeepSeek's site). 4. Track the family tree quarterly — release cycles are measured in months.

The refactor's deeper point: organizing models by vendor turns a pile of isolated feature modules into a legible map of real industrial competition — each company's technical route, product philosophy, and market position.

*Source: easy-learn-ai project — https://github.com/ConardLi/easy-learn-ai*