Restructuring a 5,000-Line AI Model Knowledge Base: The easy-learn-ai Refactor

Restructuring a 5,000-Line AI Model Knowledge Base: The easy-learn-ai Refactor

The Problem: A Single 5,000-Line File

Before July 12, 2026, the easy-learn-ai project stored its entire AI model catalog in a single file: src/utils/model.json, containing 5,005 lines of code. The data was organized alphabetically by entry — Anthropic's Claude followed by Alibaba's Qwen, then Baidu's ERNIE, then ByteDance's Doubao. DeepSeek's open-source reasoning models, Midjourney's image generators, Google's Gemini, and Kuaishou's video models all lived together in one giant JSON, making discovery difficult for developers and contributors.

The Solution: 19 Vendor-Specific Files

The refactor deleted the monolithic model.json along with its siblings img.json and video.json. In their place, 19 new JSON files were created under src/data/models/, one per AI vendor:

• deepseek.json — DeepSeek's reasoning family
• alibaba.json — Alibaba's Qwen universe
• anthropic.json — Anthropic's Claude lineup
• google.json — Google's Gemini series
• openai.json — OpenAI's GPT empire
• bytedance.json — ByteDance's Doubao and Seed
• baidu.json — Baidu's ERNIE series
• meta.json, moonshot.json, zhipu-ai.json, tencent.json, and others
• midjourney.json, pika.json, runway.json, and others for image and video generation

The 19 vendors collectively cover nearly every major player in today's AI landscape. This change upgraded the knowledge organization from a flat list to a vendor-classified taxonomy, similar to how Linnaeus organized biological species into a hierarchical system.

Key Highlights from the New Catalog

DeepSeek — Open-source reasoning evangelist. The R1 series is trained through multi-stage cold-start and reinforcement learning, enabling chain-of-thought reasoning that approaches OpenAI o1 on math, coding, and logic tasks. Six distilled descendants range from 1.5B to 70B parameters: Distill-Qwen-1.5B, Distill-Qwen-7B, Distill-Llama-8B, Distill-Qwen-14B, Distill-Qwen-32B, and Distill-Llama-70B.

Alibaba Qwen — Price-performance leader. Qwen3.5-Plus is a Mixture-of-Experts model with 397B total parameters but only 17B active per inference, beating trillion-parameter competitors on multiple benchmarks while costing 1/18 of Google Gemini 3 Pro's API price. Qwen3.7-Max offers a 1M-token context window, and Qwen3.6-Plus supports 1M context with 64K output for long-document and enterprise agent workloads.

Anthropic Claude — Long-context specialist. Claude Opus 4.8 defaults to a 1M-token context window, capable of processing an entire codebase in a single pass. Its predecessor Opus 4.6 achieved top scores on Terminal-Bench 2.0 and Humanity's Last Exam. Claude Sonnet 4.6 costs roughly half of Opus but reaches near-flagship performance on coding and computer use, scoring 72.5% on OSWorld and surpassing Opus 4.6 on financial analysis tasks.

Google Gemini — Multimodal powerhouse. The Gemini lineup spans generations 2.0 through 3.5, with Flash, Pro, and Flash-Lite tiers. Gemini 3.5 Flash exceeds Gemini 3.1 Pro on Terminal-Bench 2.1, MCP Atlas, and GDPval-AA while maintaining Flash-class speed. Gemini 2.0 Flash handles text, images, audio, video, and PDF with a 1M-token window.

ByteDance Seed — Rising contender. Seed-OSS-36B-Base is a 36B-parameter open-source model trained on 12T tokens with native 512K context, capable of processing roughly 400,000 Chinese characters at once. The doubao-seed-code and doubao-seed-2.0-code variants are optimized for code generation, completion, and refactoring.

Image and video generation. Black Forest Labs (founded by Stable Diffusion veterans) offers FLUX.1-Schnell (4-step), FLUX.1-Dev, and FLUX.1-Pro. Midjourney V7 introduces personalized models that learn user aesthetic preferences. Runway Gen-4 enables image-to-video generation with physics-aware motion. Pika 2.0/2.1 adds Pikadditions (object insertion) and Pikaswaps (element replacement) capabilities.

Structured Data Fields

Each model entry now includes the following fields:

| Field | Meaning | |-------|---------| | modelName | Official model name | | company | Owning organization | | country | Country of origin (China, USA, France, etc.) | | openSourceStatus | Open-source or proprietary | | releaseDate | Publication date | | description | Detailed plain-language description | | modelTags | Capability tags (text generation, vision, code, tool use) | | contextWindow | Maximum input length | | maxGenerationTokenLength | Maximum output length | | relatedLinks | Papers, repos, API docs | | parent | Parent model for distilled variants |

Why This Refactor Matters

1. Broader coverage — The catalog now spans 19 providers and dozens of model families, giving users a complete map of the AI landscape. 2. Clearer structure — Vendor-based organization helps users build mental models: Alibaba for price-performance, Anthropic for long context, Google for multimodality. 3. Better extensibility — Adding GPT-6, Llama-5, or a new Chinese startup's model now requires editing a single vendor file rather than navigating a 5,000-line monolith.

Context and Timeline

The AI field has accelerated dramatically: GPT-4 stunned the world in 2023, GPT-4o brought true multimodality in 2024, Claude Opus 4.1 and Gemini 2.5 Pro pushed long-context and reasoning further in 2025, and by 2026 models routinely support 1M-token contexts and outperform humans on multiple benchmarks. Projects like easy-learn-ai bridge this rapid evolution for non-expert users, providing accessible information about what frontier models can do, cannot do, and are best suited for.

---

*Based on easy-learn-ai project commit e6c189a. The project aims to make AI model knowledge accessible to everyone.*

#easy-learn-ai #daily-update #memory #kaixiao #AI-models #open-source