When an AI Knowledge Website Decided to Give Its Models Separate Homes
*Translation of a zhichai.net forum post about the Easy AI project's data refactor.*
1. An Ever-Growing "Model Encyclopedia"
Imagine organizing a library. With a few dozen books, anything goes. With thousands, finding a specific volume becomes a chore, and new librarians are afraid to touch anything.
That was roughly the situation facing the model library of the Easy AI project — a knowledge website for AI learners and developers at https://mmh1.top. Its "AI Models" module catalogs large models from GPT-4 to Claude, from Qwen to DeepSeek, with each entry containing the model name, company, open-source status, release date, capability description, feature tags, context window size, maximum output length, and related links.
All of this was stuffed into a single file: src/utils/model.json — over 5,000 lines before the refactor, plus separate image and video model files of roughly 600–700 lines each.
A 5,000-line JSON file means: scrolling to the bottom takes many mouse-wheel turns; editing a model description requires Ctrl+F through a sea of brackets, quotes, and commas; and entries from different vendors sit side by side with no clear boundaries. For maintainers, this was technical debt — one accidental deleted comma breaks the whole JSON parse. For outside contributors, the file was intimidating.
So the maintainer decided: split the models up.
2. The Art of Splitting: From Shared Courtyard to Separate Houses
On July 12, 2026, Easy AI committed a significant change (commit hash: e6c189a), summarized in one sentence:
> The 5,000-line model.json was split into 18 independent per-vendor JSON files.
The new structure (with estimated model counts):
| Vendor | File | Est. models |
|---|---|---|
| Alibaba | alibaba.json | ~15 (Qwen series) |
| Anthropic | anthropic.json | ~10 (Claude series) |
| Baidu | baidu.json | ~10 (ERNIE series) |
| Black Forest Labs | black-forest-labs.json | 2 (FLUX.2 series) |
| ByteDance | bytedance.json | ~12 (Seed / Doubao series) |
| DeepSeek | deepseek.json | ~12 (R1 series and distills) |
| Google | google.json | ~10 (Gemini series) |
| Kuaishou | kuaishou.json | ~2 |
| Meta | meta.json | ~5 (Llama series) |
| Midjourney | midjourney.json | ~1 |
| MiniMax | minimax.json | ~5 |
| Moonshot | moonshot.json | ~8 (Kimi series) |
| OpenAI | openai.json | ~20 (GPT series) |
| Pika | pika.json | ~1 |
| Runway | runway.json | ~2 |
| Stability AI | stability-ai.json | ~4 |
| Tencent | tencent.json | ~8 (Hunyuan series) |
| xAI | xai.json | ~5 (Grok series) |
| Zhipu AI | zhipu-ai.json | ~10 (GLM / ChatGLM series) |
The three original files — model.json (5,000+ lines), img.json (667 lines), video.json (491 lines) — were removed entirely, and the data-loading logic in src/utils/modelApi.ts was updated accordingly.
What does the split buy?
1. Lower maintenance cost. Updating Alibaba's Qwen entries? Just open alibaba.json — every file is only a few hundred lines.
2. Better collaboration. Two developers can edit different vendors' files simultaneously without Git merge conflicts.
3. Faster queries. The frontend can load only the vendor files it needs (e.g., all Chinese vendors) instead of parsing one giant JSON.
4. Easier extension. A new AI startup tomorrow? Add one newcompany.json without touching anything else.
This looks like a routine refactor, but it embodies an important engineering principle: once data volume passes a threshold, organization matters more than the data itself.
3. New Arrivals: Seven Vendors' Model Debuts
The refactor wasn't just moving furniture — it added substantial new content, with 7 vendors appearing for the first time:
1. Alibaba: The Qwen Empire's Dual Open/Closed Track
Alibaba's Qwen series is among the richest new content, from the open-source Qwen3.5-Plus (397B total parameters, 17B active) to the closed-source Qwen3.7-Max and scenario-specific Flash variants. The strategy is clear: build ecosystem influence with open models, serve paying customers with closed flagships. One telling detail: Qwen3.5-Plus is priced at "only 1/18 of Gemini 3 Pro's API price" — a blunt statement of the cost-effectiveness weapon in China's LLM battlefield.
2. Anthropic: The Claude Family's "Thinking Depth"
The Claude lineup spans from Haiku 4.5 to flagship Opus 4.8, which supports a 1M-token context window — roughly 750,000 characters, about two copies of *Dream of the Red Chamber*. Notably, Anthropic's "adaptive thinking" concept lets the model automatically adjust reasoning depth to task difficulty: quick answers for simple questions, deep reasoning for complex ones.
3. Baidu: ERNIE's "Knowledge-Enhanced" Route
Baidu's ERNIE series pursues a knowledge-enhanced route — combining massive text training with structured information like knowledge graphs. The newly included models cover multiple tiers from lightweight to flagship, reflecting Baidu's long accumulation in Chinese NLP.
4. Black Forest Labs: Germany's Rising Image-Generation Force
Black Forest Labs, founded by the original Stable Diffusion team, has gained traction with the FLUX.2 series. FLUX.2 [pro] and [flex] target production applications and fine-grained control respectively, with up to 4-megapixel (4MP) output. FLUX.2 uses a "Rectified Flow Transformer" architecture and integrates the Mistral-3 24B vision-language model. A technical highlight: JSON-structured control — programmatic control over composition, element placement, and style rather than relying purely on natural-language prompts.
5. ByteDance: The Seed Series' Open-Source Sincerity
The standout is Seed-OSS-36B-Base, a genuinely open model using RoPE positional encoding, GQA attention, RMSNorm, and SwiGLU activation, trained on ~12T tokens with native 512K context — roughly 380,000 characters, enough to ingest the entire first book of *The Three-Body Problem* in one go. The Doubao series also updates to Seed 2.0, spanning a general model (pro), a code model, and a multimodal understanding model (1.6).
6. DeepSeek: The Reasoning Models' "Distillation Revolution"
DeepSeek's R1 reasoning model rivals OpenAI's o1 — and it's open source, with multiple distilled versions. The newly included R1 family (from 1.5B to 70B parameters) forms a complete product matrix. Distillation lets a large "teacher" model transfer reasoning ability to smaller "student" models — meaning SMEs and individual developers can run a 7B or 14B distilled model on a laptop and get near-frontier reasoning. This is DeepSeek's biggest contribution to AI democratization.
7. Google: Gemini's "Natively Multimodal" Philosophy
Gemini is trained on mixed text, image, audio, and video data from the start, rather than bolting image understanding onto a text model. The newly included Gemini 3.5 Flash claims that on multiple coding and agentic benchmarks it outperforms the previous generation's Gemini 3.1 Pro while keeping Flash-series speed — small models can now trade blows with large ones, given sufficiently advanced architecture and training.
4. Three Industry Trends Visible in the Data
Trend 1: The Context-Window Arms Race
In 2023, GPT-4's 8K context was astonishing. By 2026, 1M-token context windows are flagship standard — Qwen3.7-Max, Gemini 2.0 Flash, and Claude Opus 4.8 all support them. Bigger contexts change application paradigms: whole-document analysis (entire legal contracts, medical reports), whole-codebase comprehension, and coherent memory across very long conversations.
Trend 2: The Open vs. Closed "Ecosystem Game"
A clear divide: open camp — Alibaba (Qwen), DeepSeek (R1), Meta (Llama), Black Forest Labs (FLUX), ByteDance (Seed-OSS), Zhipu AI (ChatGLM); closed camp — OpenAI (GPT), Anthropic (Claude), Google (Gemini), Baidu (ERNIE), ByteDance's closed Doubao series. Notably, the same company often walks both paths: Alibaba and ByteDance each open-source a lightweight/base version while closing a flagship — building developer-community influence while monetizing closed APIs.
Trend 3: From "Text Models" to Agent-Native
A high-frequency tag across these models: tool calling. Early LLMs were text generators; today's flagships can call search engines, execute code, query databases, send email. Claude Opus 4.8's adaptive thinking, Gemini 3.5 Flash's agentic-coding optimization, Qwen3.7-Max's improved agent capabilities — all point the same direction: models are evolving from "talking" to "doing."
5. Why This Matters
One might ask: what's so impressive about a knowledge site organizing model data?
Answer: information organization is itself value creation. Information overload is a real problem in AI — new models launch daily, each with a pile of technical specs, capability claims, and marketing language. Easy AI is building a structured knowledge graph of the model world: uniform entry format, vendor-based categorization for easy comparison, related links, and search/filter support. It lowers the industry's cognitive threshold — a high schooler wondering "what AI models exist" can get a full landscape from one library instead of combing through every company's tech blog.
6. Closing Thoughts
The committer behind this change works at ByteDance (lishiqi.conard@bytedance.com), but the contribution serves the whole AI community — whether you use Qwen, GPT, or open-source Llama. It points to a bigger theme: in an era of explosive AI growth, we need more knowledge infrastructure like this — not chasing the hottest news, but patiently organizing, classifying, and connecting so knowledge flows more smoothly.
> "Make learning AI truly simple" — Easy AI's slogan, and a much-needed effort of our times.
References:
- Easy AI site: https://mmh1.top
- GitHub: https://github.com/ConardLi/easy-learn-ai
- Commit: https://github.com/ConardLi/easy-learn-ai/commit/e6c189a