From a 5,000-Line JSON to 18 Vendor Files: Mapping the World of AI Models with easy-learn-ai

Forum topic · 小凯 · 2026-08-31

Summary

The open-source project easy-learn-ai restructured its AI model catalog from a single JSON file of over 5,000 lines into 18 vendor-specific files, covering OpenAI, Google, Anthropic, Meta, xAI, and Chinese companies including Alibaba (Qwen), DeepSeek, ByteDance (Doubao/Seed), Baidu (ERNIE), Moonshot AI (Kimi), MiniMax, Zhipu AI, Tencent, and Kuaishou. The refactor makes it easier to track model families and lineages such as DeepSeek-R1 distills and the GPT-4-to-GPT-5 timeline. Beyond the technical cleanup, the post highlights key industry trends: AI models are tiering into flagship, cost-effective, lightweight, and domain-specific classes; open-source models (Llama 4, DeepSeek-R1, Qwen3.5) are closing the gap with closed models; context windows have grown from 8K to 1M+ tokens; and Chinese vendors are leading in long context, open-source adoption, pricing, and multimodal capabilities. The post argues that reliable information organization is becoming as valuable as the models themselves, helping users and developers choose the right AI model among hundreds of options.

From a 5,000-Line JSON to 18 Vendor Files: Mapping the World of AI Models

*Source: easy-learn-ai commit e6c189a*

Background

In 2023, keeping track of large language models was easy — a single page could list them all. By 2025–2026, hundreds of models from dozens of companies compete for attention, with confusing naming schemes (GPT-4o, GPT-5.1, Claude 4.5, Qwen3.5, Gemini Flash...). The open-source easy-learn-ai project, an AI learning site for beginners, originally stored its entire model catalog in one JSON file of over 5,000 lines — hard to update, with no structure for model "family" relationships (e.g., DeepSeek-R1 and its distills).

The maintainer recently split this monolith into 18 independent files, one per AI vendor — US side: OpenAI, Google, Anthropic, Meta, xAI, Midjourney, Runway, Pika, Stability AI, Black Forest Labs; China side: Alibaba, Baidu, ByteDance, DeepSeek, Moonshot AI, MiniMax, Zhipu AI, Tencent, Kuaishou — totaling over 6,000 lines of data.

Key points

Model highlights by vendor

OpenAI: GPT-4 → GPT-4 Turbo → GPT-4o ("omni", first true multimodal model) → GPT-4.1 → GPT-5 → GPT-5.1. GPT-4o mini targets cheap, high-volume tasks.
Anthropic: Claude tiering named after literary forms — Opus (flagship, complex coding/research), Sonnet (best value), Haiku (fastest). Latest Claude Opus 4.8 handles 1M-character contexts; Sonnet 4.6 reportedly beats GPT-5.2 on coding benchmarks.
Google: Gemini's huge context windows (Gemini 2.0/3.5 Flash support 1M+ tokens), plus unique ecosystem integration with Search, Docs, and Gmail.
Meta: Llama open weights let small teams run near-GPT-4-level models locally. Llama 4 Maverick uses a Mixture-of-Experts architecture (400B total, 17B active); Llama 4 Scout reaches a 10M-token context window.
DeepSeek: R1 matches OpenAI o1 in math/code reasoning at a fraction of the training cost; its reasoning is distilled into small models (1.5B–70B), all fully open source. A 1.5B distill reportedly beats GPT-4o on math competitions.
Alibaba (Qwen): Qwen3.5-Plus has 397B total/17B active parameters, outperforming trillion-parameter rivals at 1/18 the price of Gemini 3 Pro; Qwen3.7-Max supports 1M-character context and 64K output.
Moonshot AI (Kimi): doubled down on long context, from 200K up to 2M characters — strong in legal document analysis and research.
ByteDance: Doubao/Seed models are highly "engineering-oriented", multimodal (text/image/video) with tool use; open-source Seed-OSS-36B supports 512K context, backed by Douyin/Toutiao data.
Baidu (ERNIE): knowledge-graph-enhanced training reduces hallucination; ERNIE-5.1-Thinking supports 128K context and output, integrated with Baidu Search.

Industry trends the refactor reveals

1. Model tiering: flagship (Opus, GPT-5, Qwen3.7-Max) → cost-effective (Sonnet, Gemini Flash) → lightweight (Haiku, DeepSeek-Distill-1.5B) → domain-specific (coding models). Choose by fit, not price. 2. Open vs. closed convergence: by 2026, open models (DeepSeek-R1, Llama 4, Qwen3.5) match or beat many closed models; enterprises increasingly adopt hybrid strategies — local open models for sensitive data, cloud closed models otherwise. 3. Context window arms race: from GPT-4's 8K tokens (2023) to 1M+ tokens (2026), enabling whole-book reading, full codebase analysis, and hours of meeting transcripts — shifting AI from Q&A chat to deep collaboration. 4. Rise of Chinese vendors: nearly half of the 18 vendors are Chinese, leading in long context, open-source downloads (Qwen), aggressive pricing (DeepSeek API at ~1/50 of OpenAI's), and multimodal generation.

What it means for you

Regular users: far more choice at lower cost — free tiers like DeepSeek for daily tasks, Kimi for long documents, Doubao for image/video, Claude for coding, Gemini for Google Workspace.
Developers/enterprises: build model-selection capability across cost, security, latency, and capability rather than chasing the newest release.

Takeaway

The refactor itself is mundane file reorganization, but its value is clear: in an era of information explosion across hundreds of models and monthly release cycles, building an ordered map of the AI landscape is a real contribution — you cannot choose well if you do not know what the options are.

---

*Based on the easy-learn-ai project, commit e6c189a.*