When an AI Knowledge Site Decided to Split Its Model Database Apart
*Translation of a forum post about the Easy AI project's data refactor (commit e6c189a, July 12, 2026).*
1. A Model Encyclopedia That Kept Getting Thicker
Easy AI (https://mmh1.top) is a knowledge site for AI learners and developers. Its "AI Models" module acts as an ever-growing encyclopedia covering models from GPT-4 to Claude, Qwen to DeepSeek. Each entry includes model name, company, open-source status, release date, capability descriptions, feature tags, context window size, max 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. There were also separate image-model and video-model files with 600–700 lines each.
A 5,000-line JSON file means editing is fragile (one misplaced comma breaks parsing), vendors' models are mixed with no clear boundaries, and external contributors are scared off. For the maintainer this was real technical debt. So the decision was made: give the models their own homes.
2. The Art of Splitting: From One Big Courtyard to Separate Houses
On July 12, 2026, Easy AI landed commit e6c189a:
> The 5,000-line model.json was split into 18 vendor-specific JSON files.
| Vendor | File | Estimated model count |
|---|---|---|
| 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 old files — model.json (5,000+ lines), img.json (667 lines), video.json (491 lines) — were fully removed, and the data-loading logic in src/utils/modelApi.ts was updated accordingly.
Benefits of the split:
- Lower maintenance cost. Updating Qwen? Just open
alibaba.json— each file is a few hundred lines. - Better collaboration. Two developers can edit different vendors without Git merge conflicts.
- Faster queries. The frontend can load only the vendor files it needs instead of parsing one giant JSON.
- Easier extension. A new company means one new
newcompany.json, touching nothing else. - Easy AI: https://mmh1.top
- GitHub: https://github.com/ConardLi/easy-learn-ai
- Commit: https://github.com/ConardLi/easy-learn-ai/commit/e6c189a
The underlying engineering lesson: once data volume passes a threshold, organization matters more than the data itself.
3. New Arrivals: Seven Vendors' Model Debuts
The refactor also added substantial new content, including first-time coverage of seven vendors:
1. Alibaba — The Qwen lineup spans open-source Qwen3.5-Plus (397B total / 17B active parameters) to closed-source Qwen3.7-Max and Flash variants. Strategy: open models for ecosystem influence, closed flagships for paying customers. Qwen3.5-Plus is pitched as "API price only 1/18 of Gemini 3 Pro." 2. Anthropic — Full Claude line from Haiku 4.5 to flagship Opus 4.8, which supports a 1M-token context window (~750k characters) and "adaptive thinking": reasoning depth adjusts automatically to task difficulty. 3. Baidu — ERNIE models follow a knowledge-enhanced路线 (knowledge-graph-augmented training), spanning lightweight to flagship tiers. 4. Black Forest Labs — FLUX.2 [pro] and [flex] image models, up to 4MP output, using a Rectified Flow Transformer architecture with a Mistral-3 24B vision-language model, and notably supporting JSON-structured control of composition and style. 5. ByteDance — Seed-OSS-36B-Base is a true open-source model (RoPE, GQA, RMSNorm, SwiGLU, ~12T training tokens, native 512K context). The closed Doubao line also gained Seed 2.0 variants (pro, code, 1.6 multimodal). 6. DeepSeek — DeepSeek-R1 and distilled variants (1.5B to 70B) form a complete matrix, letting individuals run near-frontier reasoning models on laptops — a major contribution to AI democratization. 7. Google — Gemini 3.5 Flash claims to beat the previous Gemini 3.1 Pro on several coding and agentic benchmarks while keeping Flash-class speed, reflecting the natively-multimodal, small-model philosophy.
4. Three Industry Trends in the Data
Trend 1: The context-window arms race. From GPT-4's 8K in 2023 to 1M tokens as the flagship standard in 2026 (Qwen3.7-Max, Gemini 2.0 Flash, Claude Opus 4.8). Bigger contexts enable whole-document analysis, whole-codebase understanding, and long multi-turn memory.
Trend 2: Open vs. closed ecosystem games. Open camp: Alibaba (Qwen), DeepSeek (R1), Meta (Llama), Black Forest Labs (FLUX), ByteDance (Seed-OSS), Zhipu AI (ChatGLM). Closed camp: OpenAI, Anthropic, Google, Baidu, ByteDance's Doubao. Notably, the same company often runs both tracks — open a base/light version, close a flagship — building developer goodwill while monetizing APIs.
Trend 3: From text models to agent-native. "Tool calling" is now ubiquitous: models search the web, run code, query databases. Claude Opus 4.8's adaptive thinking, Gemini 3.5 Flash's agentic coding, Qwen3.7-Max's agent improvements all point the same way: models are evolving from "can talk" to "can act."
5. Why This Matters
Information overload is a real problem in AI: new models ship daily, each with specs, claims, and marketing. Easy AI is effectively building a structured knowledge graph of the model landscape — uniform entry formats, vendor-based grouping for comparison, links for deeper reading, and search/filter support. That lowers the industry's cognitive barrier: a curious student can get a full landscape view without trawling every company's tech blog.
6. Closing Thoughts
The committer works at ByteDance (lishiqi.conard@bytedance.com), but the contribution serves the whole AI community — whether you use Qwen, GPT, or open-source Llama. In an era of explosive AI growth, we need more of this kind of knowledge infrastructure: patiently organizing, classifying, and connecting, not chasing the hottest news.
> "Making AI learning truly simple" — Easy AI's slogan, and a much-needed effort of our time.
References: