When the AI World Went From One Encyclopedia to Twenty Libraries
*Translated from a Chinese forum post on zhichai.net.*
Imagine walking into a study. For the past three years, it contained only one heavy encyclopedia—five to six thousand pages densely recording every AI model in the world. Every time you wanted to look something up, you had to open this tome and search through its vast pages.
The book itself was a miracle. It recorded the full lineage from GPT-3 to GPT-5.5, the iteration path from Claude 3.5 to Opus 4.8, and the rise of DeepSeek from V3 to V4-Pro. But the book kept getting thicker and heavier—too thick for one person to carry, too heavy to update without jostling other chapters.
Then one day, someone did something that looked small but was deeply meaningful: they split the encyclopedia into twenty independent libraries.
Why Split It?
The project's latest commit did something very programmer-like: it split model data previously crammed into one file into separate archives organized by vendor.
The old structure:
- One file with 5,000+ lines
- OpenAI's GPT series, Anthropic's Claude series, Google's Gemini series all mixed together
- Text models, image models, video models—all piled up
- One file for Tongyi Qwen's family tree
- One for Claude's evolution
- One for DeepSeek's rise
- One for the chronicle of the GPT empire
- One for Gemini and DeepMind's territory
- …20 vendors in total, each with its own home
- DeepSeek LLM (Jan 2024) — debut, 7B and 67B variants
- DeepSeek-V2 (May 2024) — MoE architecture, 236B total parameters with only 21B active
- DeepSeek-V3 (Dec 2024) — 671B total, only 37B active per token
- DeepSeek-R1 (Jan 2025) — first reasoning model, using reinforcement learning to think before speaking
- DeepSeek-V3.1/V3.2-Exp (2025) — dual thinking/non-thinking modes
- DeepSeek-V4-Pro (Apr 2026) — 1.6T total parameters, 49B active, 1M context window
- Haiku — light cavalry: fast, cheap, sufficient
- Sonnet — main force: balancing performance and cost
- Opus — heavy tank: strongest reasoning, highest price
- GPT-4o — omni-modal: text, audio, images
- o1/o3 — reasoning-focused models trained with RL to think before answering
- GPT-5 — back to a unified architecture with mini/nano/pro/codex tiers
- GPT-5.5 — 1.05M context window, 128K output
- Imagen 4 — text-to-image, up to 2K, 10x speed mode
- Nano Banana series — image editing, from consumer to 4K professional
- Veo 3/3.1 — video generation, first with native audio
- Seed-OSS-36B — open-source LLM with 512K context
- doubao-seed — two generations (1.6/2.0), covering pro/lite/flash/thinking/code
- Seedream — text-to-image, 2.0 to 4.0, 4K output
- Seedance — video generation, 1.0 to 2.0, multi-shot narrative
- Black Forest Labs (FLUX) — founded by ex-Stable Diffusion team, a new text-to-image force
- xAI (Grok) — Musk's brand, real-time data as its differentiator
- Moonshot (Kimi) — pioneer of long-context models in China, 2M context as its signature
- MiniMax — strong multimodal and voice capabilities, behind Hailuo AI
- Runway/Pika — video-generation specialists, more focused than big labs' general models
- easy-learn-ai project commit
- Official vendor documentation and model cards
Now:
On the surface, this is just code refactoring. In essence, it is an acknowledgment and mapping of the AI industry's reality.
What's Hidden in the Twenty Libraries?
The DeepSeek Pavilion: The Open-Source Rebel
A clear iteration path:
Notably, DeepSeek's distilled model matrix—R1-Distill-Qwen-1.5B/7B/14B/32B and R1-Distill-Llama-8B/70B—transfers R1's reasoning patterns to smaller models using 800K reasoning samples. And DeepSeek-Math-V2 reached gold-medal level at IMO 2025 and scored 118/120 on Putnam 2024—AI independently proving, not just assisting calculation.
The Anthropic Pavilion: Precision Watches of the Closed World
Claude's product line is a trained army:
The cadence is remarkably regular: 3.5 series (Jun–Oct 2024) introduced Artifacts; 3.7 Sonnet (Feb 2025) was the first hybrid reasoning model; the 4 series (May 2025) marked a new architectural generation; 4.1/4.5/4.6/4.8 upgraded every two to three months.
Notably, Claude Sonnet 4.6 scored highest on the Humanity's Last Exam multi-disciplinary reasoning test and led GPT-5.2 by roughly 144 Elo on GDPval-AA knowledge-work evaluation.
The OpenAI Pavilion: From the Only One to One of Many
From GPT-4 (Mar 2023) to GPT-5.5 (Apr 2026), OpenAI shows full-stack ambition:
The landmark: GPT OSS—OpenAI finally open-sourced (Apache 2.0 license). Two MoE models, 120B and 20B, 4-bit quantization, compatible with Transformers, vLLM, llama.cpp, and Ollama. Even the flag-bearer of the closed camp extended an olive branch to open source.
The Google Pavilion: The Natively Multimodal All-Rounder
From Gemini 2.0 Flash to 3.5 Flash, Google keeps doing one thing—making models that natively understand text, images, audio, video, and PDF rather than bolting capabilities on later.
Its media-generation matrix:
The Alibaba Pavilion: China-Speed Open Source
Tongyi Qwen's archive is striking—numerous, fast-iterating, aggressively open. The Qwen3 matrix includes Qwen3-235B-A22B (thinking/non-thinking), Qwen3-30B-A3B, Qwen3-4B for edge deployment, Qwen3-VL, Qwen3-Coder-480B-A35B, and the reasoning-focused QwQ-32B.
Even more aggressive: Qwen3.5-Plus—an open-source model with 397B total / 17B active parameters, API-priced at just 1/18 of Gemini 3 Pro. Plus Wan2.2 video generation, Qwen-Image text-to-image, and Z-Image (6B parameters, 8-step generation).
The ByteDance Pavilion: The Underrated All-Rounder
Especially notable: doubao-seed-1.6-thinking, a thinking-mode model with 32K chain-of-thought, strongly enhanced in math, coding, and logical reasoning.
Five Industry Codes in the Data
1. Big Models Are No Longer a Few Players' Game
20 vendors—unimaginable three years ago, when the club was just OpenAI, Google, and Anthropic. Chinese vendors (Alibaba, ByteDance, Baidu, DeepSeek, Moonshot, MiniMax, Zhipu, Tencent, Kuaishou) now occupy half the map.
2. MoE Has Become the Mainstream Choice
DeepSeek-V4 (1.6T/49B), Qwen3.5-Plus (397B/17B), GPT OSS (120B MoE)—nearly every new model uses Mixture-of-Experts. The logic: bigger models are more capable, but inference costs must stay controllable. MoE maintains a vast expert pool but only wakes the relevant few per query.
3. Context Windows Are in an Arms Race
64K → 128K → 256K → 1M → 1.05M. GPT-5.5: 1,050K. Gemini 3.5 Flash: 1,048,576 tokens. DeepSeek-V4-Pro: 1M. Context growth outpaces Moore's law—AI is moving from short-term to long-term memory, fitting whole books, codebases, or a year of email in one pass.
4. The Awakening of Reasoning
From DeepSeek-R1 to OpenAI o1/o3, from Claude's hybrid reasoning to Qwen's thinking mode—teaching AI to think before speaking is industry consensus. This is a paradigm shift: from predicting the next token to training chains of thought via reinforcement learning.
5. The Open vs. Closed Tug-of-War
Open-source camp: DeepSeek (all open), Alibaba Qwen (mostly open), Meta Llama, Google Gemma, OpenAI GPT OSS. Closed camp: GPT-5.5, Claude, Gemini. Open models are rapidly closing the performance gap; DeepSeek-V4-Pro is officially positioned as a high-capability open model approaching top closed models.
One Refactor, One Mirror
Why split? The answer is simple: because the world changed. Three years ago, one encyclopedia sufficed. Today—20 vendors, hundreds of models spanning text/image/video/audio/code—no single file can hold that complexity.
This refactor isn't just paying off technical debt; it's an honest mapping of industry reality. AI is no longer one company's solo act but a multi-polar, layered, ecosystem-like world. Each vendor's JSON file is a library; together they form a living map of the AI industry—updated daily.
Easter Eggs for the Curious
> The number of models will explode, but the framework for understanding them must stay simple. > > The biggest lesson: in an age of information overload, organizational structure is itself information. Splitting 5,000 lines into 20 files isn't just tidying data—it's helping everyone understand the world.
References