When AI Learns to Sort Itself: The Map of the Model World Behind a Project Refactor

Forum topic · 小凯 · 2026-09-05

Summary

A refactor of the open-source easy-learn-ai project split a single 5,005-line model.json file into 20 per-vendor JSON files under text, image, and video directories, following the separation-of-concerns principle. The reorganized database now clearly reveals the landscape of 240 AI models from 20 companies: OpenAI leads with 38 models spanning every price tier; 9 Chinese vendors (Alibaba, Zhipu AI, ByteDance, DeepSeek, Moonshot, Baidu, Tencent, MiniMax, Kuaishou) contribute over a third of all models; DeepSeek has 11 of its 17 models as open-source distills of R1, enabling small models like Qwen-1.5B to approach GPT-4-level math reasoning; Anthropic bets on long context windows up to 1M tokens; image models (Midjourney, FLUX, Stability AI) and video models (Runway, Pika, Kuaishou Kling) mark new battlegrounds; and the open-source versus closed-source split shows Meta, Alibaba, DeepSeek, and Zhipu AI promoting open weights while OpenAI, Anthropic, Google, and most Chinese closed-model vendors keep weights private. The refactor turned routine code cleanup into a snapshot of the entire AI industry.

Key points

This post chronicles a refactoring of easy-learn-ai, an open-source website helping ordinary people learn about AI models, and reflects on what the reorganized data reveals about the AI industry.

The refactor

The project's entire model database was crammed into a single model.json file of 5,005 lines, mixing text, image, and video models from all vendors.
It was split into 20 per-vendor JSON files (e.g., alibaba.json, anthropic.json, deepseek.json, openai.json, zhipu-ai.json), with image models moved to img/ and video models to video/ directories.
This applies the principle of Separation of Concerns: easier searching, safer edits, fewer Git merge conflicts.

The data: 240 models from 20 vendors

OpenAI leads with 38 models, covering every price and capability tier (GPT-5, GPT-5 mini, GPT-4o mini, GPT-4.1 nano, GPT-4 Turbo, etc.).
9 of 20 vendors are Chinese, together accounting for over a third of all models:
Alibaba (28 models): Qwen series, aggressively open-source, e.g., a 397B-parameter MoE (Qwen3.5-Plus) claimed to outperform trillion-parameter closed models at 1/18 the price of Gemini 3 Pro.
Zhipu AI (24 models): ChatGLM/GLM series, early open-source Chinese LLMs.
ByteDance (20 models): Seed series, powering Douyin AI features.
DeepSeek (17 models): 11 of them are distilled from DeepSeek-R1, which generated 800K reasoning samples from its 671B-parameter model; a 1.5B Qwen distill reportedly rivals GPT-4 in math reasoning. All distills are open source.
Anthropic (12 models) bets on long context: Claude Opus 4.8 supports 1M tokens (~750K Chinese characters).
Image/video battlefield: Midjourney (brand as category), Black Forest Labs' FLUX.1 (open-source image quality rivaling Midjourney), Runway Gen-4 and Pika 2.0 in video, plus Kuaishou's Kling showing Chinese teams are competitive in video generation.

Capability tags

The database's capability tags map AI's ability layers: text generation (universal baseline), deep thinking (chain-of-thought, e.g., DeepSeek-R1, Claude Opus), visual understanding (GPT-4o, Claude, Gemini, Qwen), tool calling (agentic behavior, with Claude Sonnet 4.6 furthest along in computer use), and code generation (competed via SWE-bench scores).

Open vs. closed source

Closed: OpenAI, Anthropic, Google, Baidu (ERNIE), ByteDance (Seed), Moonshot (Kimi), MiniMax, Tencent (Hunyuan).
Open: Meta (Llama), Alibaba (Qwen), DeepSeek, Zhipu AI, Stability AI, Black Forest Labs (FLUX).
Open vendors prioritize ecosystem and brand reach; closed vendors protect core IP. Open source democratizes AI by giving developers control over their own data and models.

Context windows

Context length has grown from GPT-3.5's 4K tokens to 128K–1M in current models — the key upgrade that turns AI from a chat tool into a work partner that can ingest entire books, a year of emails, or full financial records.

Conclusion

What began as routine code cleanup ended up drawing a map of the AI world: US–China competition, the open/closed divide, new image and video fronts, the context-length arms race, and distillation decoupling capability from parameter count.

---

Data source: easy-learn-ai project commit e6c189a — 20 vendors, 240 models, covering OpenAI, Anthropic, Google, Meta, xAI, Stability AI, Midjourney, Runway, Pika Labs, Black Forest Labs, Alibaba, Baidu, ByteDance, DeepSeek, Zhipu AI, Moonshot AI, MiniMax, Tencent, and Kuaishou.