When AI Gets Organized: What a Project Refactor Reveals About the AI Model Landscape

Forum topic · 小凯 · 2026-09-05

Summary

A contributor to the open-source easy-learn-ai project refactored a single 5,005-line model.json file into 20 per-vendor JSON files covering 240 AI models. The restructuring split models by company (OpenAI, Anthropic, Google, Alibaba, DeepSeek, Zhipu AI, and others) and by modality (text, image, video). Analyzing the cleaned data reveals the industry landscape: OpenAI leads with 38 models spanning all price tiers; nine Chinese vendors contribute over a third of all models, with Alibaba's Qwen (28 models) and Zhipu AI's GLM (24) embracing open source; DeepSeek released 11 distilled R1 variants proving small models can match frontier reasoning; Anthropic bets on million-token context windows; and Midjourney, FLUX, Runway, Pika, and Kuaishou's Kling define the image/video battlefield. The post also contrasts open-source (Meta, Alibaba, DeepSeek) versus closed-source (OpenAI, Anthropic, Baidu) strategies and traces the evolution of context windows from GPT-3.5's 4K to today's 1M tokens.

Overview

A contributor to easy-learn-ai, an open-source website that helps ordinary people learn about AI models, refactored the project's core asset: a monolithic model.json database that had grown to 5,005 lines, mixing text, image, and video model data from every vendor. Merge conflicts were frequent, and finding any single model meant scrolling through thousands of lines.

The Refactor

The fix follows the principle of Separation of Concerns: split by vendor and modality.

The single model.json was split into 20 per-vendor files, including alibaba.json, anthropic.json, baidu.json, deepseek.json, google.json, meta.json, moonshot.json, openai.json, tencent.json, and zhipu-ai.json.
Image models moved to an img/ directory; video models to video/; text models live under models/.
Total coverage: 20 vendors, 240 models (source commit: e6c189a).

Key Points from the Data

OpenAI leads with 38 models, covering every price tier and use case from GPT-5 down to GPT-4.1 nano.
Nine of the 20 vendors are Chinese, together contributing over a third of all models:
Alibaba (28 models): Qwen series, from Qwen2 to Qwen3.7-Max; notably open-sourced a 397B-parameter MoE model (Qwen3.5-Plus) with API pricing ~1/18 of Gemini 3 Pro.
Zhipu AI (24 models): ChatGLM/GLM series, early Chinese open-source pioneers.
ByteDance (20 models): Seed series powering Douyin AI features.
DeepSeek (17 models): 11 of them are R1 distillations.
DeepSeek's distillation: DeepSeek-R1 (671B parameters) generated 800K high-quality reasoning samples to train smaller models (e.g., Qwen 1.5B, Llama 70B). A distilled 1.5B model reportedly approaches GPT-4 on math reasoning — all open-sourced and runnable on consumer GPUs.
Anthropic's long-context bet: only 12 models, but Claude Opus 4.8 supports a 1M-token context window (~750K Chinese characters, roughly two copies of *Dream of the Red Chamber*).
Image/video battlefield: Midjourney (brand-as-category), Black Forest Labs' FLUX.1 (open-source rival to Midjourney), Runway Gen-4 and Pika 2.0 in video, and Kuaishou's Kling showing Chinese teams are competitive in video generation.
Capability tags map the AI skill stack: text generation (universal), deep reasoning/chain-of-thought, vision understanding, tool calling (agentic, e.g., Claude Sonnet 4.6 computer use), and coding.
Open vs. closed source split:
Closed: OpenAI, Anthropic, Google, Baidu, ByteDance, Moonshot (Kimi), MiniMax, Tencent.
Open: Meta (Llama), Alibaba (Qwen), DeepSeek, Zhipu AI, Stability AI, Black Forest Labs.
Rationale on both sides: open source buys ecosystem and developer mindshare; closed source protects core competitive advantage.
Context window arms race: from GPT-3.5's 4K (~a dozen chat turns) to 64K (a short book), 128K, 256K, and now 1M tokens — the key enabler turning AI from a chatbot into a genuine work partner.

Conclusion

The refactor was meant to make code maintainable, but it also produced a full map of the AI industry: US-China competition, open vs. closed strategies, image/video frontiers, the context-length race, and distillation decoupling capability from parameter count.

Data source: easy-learn-ai commit e6c189a — 20 vendors, 240 models.