English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

easy-learn-ai Restructures Its AI Model Knowledge Base: From a 5,000-Line JSON to 19 Vendor Files

Forum topic · 小凯 · 2026-08-11

Summary

The open-source project easy-learn-ai replaced its monolithic 5,005-line model.json (along with img.json and video.json) with 19 vendor-specific JSON files under src/data/models/, covering DeepSeek, Alibaba, Anthropic, Google, OpenAI, ByteDance, Baidu, Meta, Moonshot, Zhipu AI, Tencent, Black Forest Labs, Midjourney, Runway, Pika, and more. The post walks through highlights: DeepSeek-R1 and its six distilled models (1.5B-70B); Alibaba's Qwen3.5-Plus MoE with 397B total/17B active parameters priced at 1/18 of Gemini 3 Pro, plus 1M-token Qwen3.7-Max; Claude Opus 4.8 with a default 1M-token context; Gemini 3.5 Flash beating Gemini 3.1 Pro on coding benchmarks; ByteDance's open Seed-OSS-36B with 512K context; and image/video generators like FLUX.1, Midjourney V7, Runway Gen-4, and Pika 2.x. Each entry now carries structured fields such as open-source status, release date, context window, tags, and parent model, improving discoverability, clarity, and maintainability. Source: commit e6c189a on GitHub.

Background

Before July 12, 2026, the easy-learn-ai project stored all of its AI model information in a single file, src/utils/model.json — 5,005 lines of code covering models from every major vendor in alphabetical order, with no categorization by company, capability, or release date. This post describes a major restructuring of that knowledge base.

What Changed

The original model.json, along with img.json and video.json, was deleted and replaced by 19 new JSON files under src/data/models/, one per vendor:

  • deepseek.json — DeepSeek's reasoning family
  • alibaba.json — Alibaba's Qwen lineup
  • anthropic.json — Anthropic's Claude models
  • google.json — Google's Gemini series
  • openai.json — OpenAI's GPT family
  • bytedance.json — ByteDance's Doubao and Seed models
  • baidu.json — Baidu's ERNIE series
  • Plus meta.json, moonshot.json, zhipu-ai.json, tencent.json, and image/video vendors such as midjourney.json, pika.json, and runway.json.
  • Highlights from the New Data

    DeepSeek

  • R1: a reasoning model trained via multi-stage cold start and reinforcement learning, approaching OpenAI o1 on math, coding, and logic; distilled into six smaller models: Distill-Qwen-1.5B, Distill-Qwen-7B, Distill-Llama-8B, Distill-Qwen-14B, Distill-Qwen-32B, Distill-Llama-70B.
  • Alibaba Qwen

  • Qwen3.5-Plus: a Mixture-of-Experts model with 397B total parameters but only 17B active, claimed to beat trillion-parameter rivals on several benchmarks at ~1/18 the API cost of Google Gemini 3 Pro.
  • Qwen3.7-Max: 1M-token context window; Qwen3.6-Plus: 1M context and 64K output for long-document and enterprise agent use.
  • Anthropic Claude

  • Claude Opus 4.8: default 1M-token context; top scores cited on Terminal-Bench 2.0 (coding) and Humanity's Last Exam (multi-disciplinary reasoning).
  • Claude Sonnet 4.6: roughly half the price of Opus, "near-flagship" on coding and long-context reasoning; 72.5% on OSWorld computer use, and reportedly beats Opus 4.6 on financial analysis tasks.
  • Google Gemini

  • Gemini 3.5 Flash: per Google, surpasses Gemini 3.1 Pro on Terminal-Bench 2.1, MCP Atlas, and GDPval-AA while keeping Flash-tier speed.
  • Gemini 2.0 Flash: multimodal input (text, image, audio, video, PDF) with a 1M-token context window.
  • ByteDance Seed

  • Seed-OSS-36B-Base: 36B parameters, 12T training tokens, native 512K context — open source alongside the closed Doubao line.
  • doubao-seed-code / doubao-seed-2.0-code: coding-optimized models for generation, completion, refactoring, and project-level code understanding.
  • Image and Video Generation

  • Black Forest Labs FLUX.1: Schnell (4-step fast), Dev (high quality), Pro (API).
  • Midjourney V7: strong aesthetics and prompt understanding, with personalized models that learn user preferences.
  • Runway Gen-4: image-to-video generation with physically plausible motion.
  • Pika 2.0/2.1: Pikadditions (insert objects into existing video) and Pikaswaps (replace elements in video).

Structured Metadata

Each model entry now uses consistent fields: modelName, company, country, openSourceStatus, releaseDate, description, modelTags, contextWindow, maxGenerationTokenLength, relatedLinks, and parent (e.g., linking distilled models to their teacher).

Why It Matters

1. More complete coverage: 19 vendors and dozens of model families in one graph. 2. Clearer structure: vendor-based organization helps users build a mental map of each company's strengths. 3. Better extensibility: new models only require adding a record to the relevant vendor file, not editing a 5,000-line monolith.

---

*Based on easy-learn-ai commit e6c189a.*

Tags

#easy-learn-ai#ai-models#open-source#json-restructuring#llm#data-organization#deepseek#qwen

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633331