Giving AI Models a Family Registry: A Restructuring of Knowledge Order
*(Translation of a zhichai.net forum post about the easy-learn-ai project, https://github.com/ConardLi/easy-learn-ai, commit e6c189a)*
1. A Library in Disarray
Imagine walking into a vast library that collects not ordinary books, but humanity's wildest recent creations—AI models. From ChatGPT that writes poetry, to Midjourney that paints, from Claude that codes, to Runway that generates video—each one is a brilliant star.
The problem: these stars were piled together randomly.
Some were organized by "what they can do"—a "text generation zone" here, an "image generation zone" there, a few "video generation zones" stuffed in a corner. If you found a model that could both write articles and draw images, where would it go? Worse, to track Chinese vendors' progress you had to search three different zones; to compare OpenAI and Anthropic products of the same generation, you had to remember complex model IDs and dig through stacks.
This was the easy-learn-ai project's situation before restructuring: all model data crammed into a few giant JSON files—model.json for text models, img.json for image models, video.json for video models. Like a library sorting all books by thickness—technically correct, but maddening to use.
2. Why "By Vendor" Is Smarter
The restructuring did something seemingly simple but profound: reorganized model data from "by capability" to "by family." Each AI vendor now has its own registry:
- alibaba.json — Alibaba's Qwen family, from Qwen3-Max to Wan2.2 video models
- anthropic.json — Anthropic's Claude family, the full Opus-to-Haiku lineup
- deepseek.json — DeepSeek's R1/V3/V4 series
- google.json — Google's Gemini empire, from 2.0 Flash to Veo video generation
- openai.json — OpenAI's GPT series
- baidu.json, bytedance.json, tencent.json… and more
- Claude Opus 4.8: 1000K (1 million tokens)
- DeepSeek-V4-Pro: 1000K
- Qwen3.7-Max: 1000K
- Gemini 2.5 Pro: 1000K
- DeepSeek-V4-Pro: 384K output (enough for a novella)
- Claude Opus 4.8: 128K
- Gemini 2.5 Pro: 65K
- Most earlier models: 8K–32K
- Generalists are multiplying. Claude Opus 4.8, Gemini 3.5 Flash, and Qwen3.5-Plus carry all tags—they read images, write code, reason deeply, call tools. AI is shifting from specialists to generalists.
- "Deep thinking" is standard equipment. Once unique to DeepSeek-R1 and Claude Opus, reasoning mode is now supported by nearly every vendor's flagship—Qwen has Thinking editions, Gemini has Thinking mode, DeepSeek has V3.2-Speciale for high-reliability reasoning. The market has validated that users pay for deeper thinking.
- Multimodality is swallowing everything. Pure-text models' space is shrinking rapidly. Alibaba's Qwen3-Omni understands text, images, audio, and video, and generates speech; Google's Gemini series is natively multimodal. The future AI won't "only chat"—it will see, hear, speak, and draw.
Why is this smarter? Because AI models don't exist in isolation—they have genealogies.
Take Claude: Claude 3.5 Sonnet → Claude 4 Sonnet → Claude 4.5 Sonnet → Claude 4.6 Sonnet. A clear product line, each generation evolving from the last. Scattered across a "text zone," that lineage disappears. But lined up in anthropic.json, you see Anthropic's product strategy at a glance—flagship Opus breaks the ceiling, workhorse Sonnet balances performance and cost, lightweight Haiku pursues extreme speed.
Or DeepSeek: from early DeepSeek LLM (January 2024), to DeepSeek-V2 (a 236B-parameter MoE experiment), to the world-shaking DeepSeek-R1 (reasoning benchmarked against OpenAI-o1), then V3 (a 671B MoE monster), V3.1 (dual-mode thinking), V3.2-Exp (DSA sparse attention cutting cost 50%), V4-Pro (1.6T total parameters, 1M context), and finally V4-Flash (a 284B high-speed lightweight version). That timeline in deepseek.json is an epic of Chinese open-source AI.
3. Ambition in the Details
The new schema contains carefully designed fields, each answering a key question:
#### contextWindow: a model's "memory length"
Measured in K (thousands of tokens). Four flagship models now report 1000K:
A million tokens is roughly 750,000 Chinese characters, or three copies of *Dream of the Red Chamber*. You can feed an entire novel for character and plot analysis without the model "forgetting the beginning." The arms race is white-hot: in 2024, 128K was flagship; by 2026, 1000K became standard across vendors. Why? Because long context is the infrastructure of agents—an AI that can process million-word documents can genuinely replace lawyers reading contracts, doctors reading charts, researchers reading papers.
#### maxGenerationTokenLength: a model's "desire to express"
Output limits vary enormously:
Output length determines what a model can do: 8K suits chat; 32K, technical reports; 128K, full codebase documentation; 384K, an entire structured book in one shot. DeepSeek's aggressiveness here reflects its focus on long-form generation.
#### modelTags: a model's "skill tree"
Tags include text generation, deep thinking, visual understanding, code enhancement, tool calling, image generation, video generation. They sketch 2026's capability landscape:
This field reveals derivation. DeepSeek-R1-Distill-Qwen-32B's parent is DeepSeek-R1—it's a smaller model "distilled" from R1. Like a family tree, it explains why certain models share capabilities or suit local deployment.
4. Industry Trends Read from the Data
Across the 19 vendor files, a clear map emerges:
US-China dual-pole solidified. US camp: OpenAI, Anthropic, Google, Meta, xAI, Midjourney, Runway, Pika, Stability AI. China camp: Alibaba, DeepSeek, Baidu, ByteDance, Tencent, Moonshot, MiniMax, Zhipu AI, Kuaishou. US vendors chase closed-source flagship performance (Claude Opus 4.8, Gemini 3.5 Flash, GPT-5); Chinese vendors push open source and value (DeepSeek-V4, Qwen3.5-Plus).
"Small parameters, big wisdom" is the new paradigm. DeepSeek-V4-Pro has 1.6T total parameters but activates only 49B per query; Qwen3.5-Plus totals 397B but activates 17B. Mixture-of-Experts architecture lets large models run on consumer hardware—key to open-source adoption.
Reasoning is the core competitive capability. In 2025, DeepSeek-R1 proved via reinforcement learning that letting models "think longer" beats merely stacking parameters. By 2026, everyone followed: Claude adaptive thinking, Gemini Deep Think, Qwen Thinking mode. Slow thinking went from selling point to standard.
Multimodality went from show-off to practical. Google's Veo 3.1 generates 1080P video with sound; Alibaba's Wan2.5 supports audio-visual sync and 10-second 1080P; Runway's Gen-4 offers 60-second generation. These are entering real workflows in advertising, film, and e-commerce.
5. One Refactoring, One Cognitive Upgrade
On the surface, this was technical debt cleanup; in reality, a cognitive framework upgrade. Organizing models by vendor means seeing AI not as isolated feature modules but as a real industry competition—each vendor with its own technical路线, product philosophy, and market position. Anthropic pursues safety and alignment, OpenAI general intelligence, DeepSeek open source and value, Google multimodal integration, Alibaba full-stack coverage.
This matters for learners. Don't just ask "what is a Transformer"—ask "why did DeepSeek choose MoE," "why does Claude stick with Constitutional AI," "why did Gemini push context to 1 million tokens." No standard answers, but these questions constitute real industry understanding.
6. A Beginner's Guide
1. Pick one "family" to study deeply. Start with DeepSeek (clear R1-to-V4 evolution, many technical reports) or Claude (representative closed-source safety research). 2. Watch context window and output length. A 1000K-context + 128K-output model and a 128K-context + 8K-output model are entirely different creatures. 3. Try things hands-on. Test Qwen3-Max on DashScope, Claude Sonnet on Anthropic Console, DeepSeek V3.2 on its official site. The differences in style are worth more than a hundred papers. 4. Track the genealogies. Claude went from 3.5 to 4.8 in about eighteen months; DeepSeek from V2 to V4 in one year. This industry moves faster than you think.
7. Epilogue
After the restructuring, easy-learn-ai's model library went from a few chaotic files to 19 clear "family archives," each containing a vendor's technical ambition and product story. More stories are coming—Claude 5 may be training, DeepSeek-V5 may be en route, Gemini 4 may be imminent.
One thing is certain: AI models are becoming "understandable, comparable, selectable" infrastructure rather than black-box toys. The first step to understanding them is knowing where they come from, who owns them, and what they can do. This refactoring does exactly that.
---
*Source: easy-learn-ai project (https://github.com/ConardLi/easy-learn-ai), commit e6c189a — feat: added multiple model entries covering Alibaba, Anthropic, Baidu, Black Forest Labs, ByteDance, DeepSeek, and Google's latest models.*