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

When There Are Too Many AI Models to Remember: easy-learn-ai Draws a New Map by Splitting model.json Into Per-Brand Files

Forum topic · 小凯 · 2026-08-20

Summary

A forum post on zhichai.net reviews a refactor of the open-source easy-learn-ai project, which previously stored all LLM metadata in a single 5,000-line model.json file. The commit splits the data into 20 per-brand JSON files under src/data/models/ — covering OpenAI, Google, Anthropic, Meta, Alibaba (Qwen), DeepSeek, Moonshot (Kimi), ByteDance (Seed), Baidu, Tencent, Zhipu, xAI, MiniMax, Midjourney, Runway, Pika, Kuaishou, Stability AI, and Black Forest Labs. Beyond the restructuring, the post highlights model-lineup details and three industry trends: 1M-token context windows becoming the baseline (DeepSeek V4, Gemini 2.0 Flash, Kimi-Linear-48B, GPT-4.1/5, Qwen3.6-Plus); Mixture-of-Experts architectures becoming standard (DeepSeek V3 671B/37B active, Kimi K2 1T/32B, Qwen3.5-Plus 397B/17B); and toggleable reasoning modes (Thinking/Non-Thinking, OpenAI reasoning.effort, K2-thinking, ERNIE-5.x-Thinking). The modular per-brand structure makes ongoing maintenance sustainable as new models launch.

Introduction

The post opens with an analogy: the AI model ecosystem used to be like an old-fashioned grocery store — everything piled together, navigable only by the owner who had memorized it. The easy-learn-ai project set out to fix this with a beginner-friendly LLM knowledge base, but rapid model releases (DeepSeek V4, GPT-5.5, Kimi K2.6, etc.) caused all information to pile into a single model.json file of over 5,000 lines.

The Refactor: Splitting the Shelf

A recent commit decomposed model.json into 20 independent brand-specific JSON files under src/data/models/:

  • alibaba.json — the Qwen family
  • anthropic.json — Claude models
  • baidu.json — ERNIE models
  • black-forest-labs.json — FLUX image models
  • bytedance.json — ByteDance Seed models
  • deepseek.json — DeepSeek models
  • google.json — Gemini models
  • kuaishou.json — Kling video models
  • meta.json — Llama models
  • midjourney.json, minimax.json, moonshot.json, openai.json, pika.json, runway.json, stability-ai.json, tencent.json, xai.json, zhipu-ai.json
  • Each file records model parameters, context windows, capability tags, release dates, and links.

    Highlights From Selected Brands

    DeepSeek

  • Full R1-Distill lineup (Qwen-based 1.5B / 7B / 14B / 32B / 70B) enabling local inference with complete chain-of-thought structure.
  • V4 series (Pro and Flash): MoE with 1.6T total parameters but only 49B activated per token — cost-efficient at scale.
  • Moonshot (Kimi)

  • Evolution: moonshot-v1-8k → K2 (1T-param open-source MoE) → K2.5 (Agent Swarm coordinating up to 100 sub-agents) → K2.6 (multimodal agent model).
  • Kimi-Linear-48B uses linear attention to cut KV cache by 75% and speed up decoding 6x, with a 1M context window.
  • OpenAI

  • Longest lineup: GPT-3.5 through GPT-5.5, o1 to o4-mini, Codex series, and GPT OSS open-source releases (120B/20B, Apache 2.0).
  • GPT-5.5: 1050K context; GPT-5.5 Pro for hard problems; GPT-5.3-Codex for coding.
  • "Minimal reasoning" option in GPT-5 series lets users dial reasoning effort from none to deep.
  • Alibaba (Qwen)

  • Qwen3.5-Plus: 397B total / 17B active parameters, claimed to beat the trillion-param Qwen3-Max at 1/18 the API price of Gemini 3 Pro.
  • Full size coverage from 0.5B lightweight to 3.7-Max flagship, open and closed, text to multimodal.
  • ByteDance (Seed)

  • Seed-OSS-36B-Base released open source: 512K context, 12T training tokens, adjustable thinking budget.
  • Doubao-Seed-Code: 256K context, image/video understanding, chain-of-thought plus tool calling; plus doubao-1.5-pro/vision/lite variants.
  • Three Industry Trends

    1. 1M-token context is the new baseline — DeepSeek V4, Gemini 2.0 Flash, Kimi-Linear-48B, GPT-4.1/5, and Qwen3.6-Plus all offer 1M (roughly 750k Chinese characters, ~3 copies of Dream of the Red Chamber), enabling whole-novel analysis, full codebase review, or year-long financial trend analysis in one pass.

    2. MoE is now standard — DeepSeek V3 (671B total / 37B active), Kimi K2 (1T / 32B), Qwen3.5-Plus (397B / 17B), GPT OSS (120B). Like a hospital triaging specialists, MoE routes each input to relevant "experts."

    3. Reasoning modes are a standard knob — DeepSeek V3.2/V4 Thinking/Non-Thinking, OpenAI reasoning.effort from none to xhigh, Kimi K2-thinking, Baidu ERNIE-5.x-Thinking. Users can now trade latency for depth per task.

    Why the Map Matters

    The refactor is more than code cleanup: it makes the ecosystem navigable and the project sustainable. New models become a single entry in a brand file; new series become a new file. A 5,000-line monolith becomes a parallel-maintainable system — like a city growing from one main road into a grid.

    References

  • easy-learn-ai project commit e6c189a
  • DeepSeek API docs: https://api-docs.deepseek.com
  • Kimi platform: https://platform.moonshot.cn
  • OpenAI models docs: https://platform.openai.com/docs/models
  • Google Gemini API docs: https://ai.google.dev/gemini-api/docs/models
  • Alibaba Qwen model list: https://help.aliyun.com/zh/model-studio/text-generation-model/

Tags

#easy-learn-ai#llm#model-landscape#moe#context-window#deepseek#kimi#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/178633722