A Library Without Order
Imagine walking into a library where every book, regardless of language, era, or subject, has been stuffed into a few oversized crates. Want *The Three-Body Problem*? Start with crate A and check thousands of titles. Want *Dream of the Red Chamber*? It might be at the bottom of crate B, pinned beneath a 2023 AI paper.
That was easy-learn-ai before July 12, 2026.
The project's mission is simple: maintain a comprehensive catalog of AI model metadata so anyone studying large models has a reliable reference. The problem was that all model data, covering text, image, and video generation from OpenAI, Alibaba, obscure startups, and everyone in between, was stuffed into three files: model.json, img.json, and video.json.
Three files. Nearly six thousand lines.
Six thousand lines of JSON is like a giant, unpartitioned hard drive. Every lookup is an adventure. Maintenance is worse: when DeepSeek ships a new version or Google adds a Gemini sub-variant, the maintainer must locate the right line inside the tangle, edit it, and pray nothing else breaks.
Categorization Is a Worldview
On July 12, 2026, developer lishiqi.conard did something that looks trivial but carries deep implications: the three crates were broken apart, and not arbitrarily. They were split by vendor.
OpenAI models go to openai.json. Alibaba models go to alibaba.json. Anthropic, Baidu, ByteDance, DeepSeek, Google, and others each get their own file. In total, 19 files were produced, covering 19 AI vendors.
| Vendor file | Lines | Representative family |
|-------------|-------|-----------------------|
| openai.json | 981 | GPT series, o series |
| alibaba.json | 752 | Qwen series |
| zhipu-ai.json | 570 | ChatGLM series |
| bytedance.json | 518 | Seed series |
| deepseek.json | 487 | DeepSeek-R1/V3 series |
| google.json | 421 | Gemini series |
| moonshot.json | 365 | Kimi series |
| baidu.json | 362 | ERNIE series |
| tencent.json | 353 | Hunyuan series |
| anthropic.json | 373 | Claude series |
(The remaining vendor files are smaller, often single-domain companies.)
This is not just code refactoring. It is a rewriting of worldview.
Previously, the project asked: "What can these models do?" (sorted by capability: text, image, video). Now, the project asks: "Where do these models come from?" (sorted by vendor).
That shift matters. In AI, "who built it" is becoming more important than "what it can do."
Nineteen Kingdoms
Let's crack open a few crates.
DeepSeek: The Open-Source Disruptor
deepseek.json lists nearly 20 models, from the flagship DeepSeek-R1 to a series of distilled variants (Distill-Qwen-1.5B, 7B, 14B, 32B; Distill-Llama-8B, 70B).
A detail worth noting: each distilled model carries a parent field pointing to DeepSeek-R1. This is a family tree. R1 is the ancestor; the distilled variants are descendants with different "genes" (some inherited from Qwen, others from Llama).
DeepSeek's story is well known: a Chinese company trained a reasoning model competitive with OpenAI at a fraction of the cost, then open-sourced it. When you see these models listed line by line, with context windows ranging from 64K to 128K and an openSourceStatus field uniformly marked "open source," the impact of that decision becomes tangible.
DeepSeek didn't release a model. It released a family.
Anthropic: The Flagship of Safety
The Claude series in anthropic.json reads like a flagship phone's iteration history: Opus 4.6, Opus 4.8, Sonnet 4.6, Haiku 4.5.
Anthropic's model descriptions repeatedly emphasize: "agentic coding," "long-context stability," "tool-trigger reliability," and "honesty." This reveals the company's obsession, not just a smart AI, but a trustworthy AI.
Claude Opus 4.8 supports a one-million-token context window. That is roughly equivalent to feeding *War and Peace* into the model and asking it to summarize the main character relationships, which it can actually do.
Alibaba: The Two Faces of Open and Closed
alibaba.json may be the most "split" file of the lot.
On one side, Qwen3.5-Plus is marked "open source," with 397B total parameters, 17B activated, an MMLU-Pro score of 87.8%, and a price only 1/18 of Gemini 3 Pro. That is a generous gift to the open-source community.
On the other side, Qwen3.7-Max and Qwen3.6-Plus are marked "closed source," exclusively available on the Alibaba Cloud Model Studio (Bailian) platform, supporting a 1M-token context.
Open source builds brand. Closed source earns profit. That strategy is not new, but seeing it written so clearly in the same JSON file forces you to admire Alibaba's shrewdness in the AI era.
Google: From Search King to Model Factory
The Gemini series in google.json reveals a tech giant's anxiety and ambition.
Gemini 2.0 Flash emphasizes "high throughput and low cost," with a 1M context window and native multimodality, clearly aimed at large-scale applications.
Gemini 3.1 Pro hits 77.1% on the ARC-AGI-2 benchmark, a test focused on abstract reasoning, not a leaderboard one can climb by memorization.
Gemini 3.5 Flash is even more interesting: Google claims it surpasses 3.1 Pro on multiple coding and agent benchmarks while still being labeled "Flash" (the fast tier). Google is blurring the line between flagship and fast food. Soon, "fastest" and "strongest" may be the same thing.
Context Window: A Silent Arms Race
Scan the contextWindow field of every model and a striking trend emerges.
GPT-4 (2023): 8K. GPT-4o (2024): 128K. Gemini 2.0 Flash (2025): 1M (1000K). Claude Opus 4.8 (2026): 1M. Qwen3.7-Max (2026): 1M.
The context window determines how much text an AI can "read" at once.
8K is roughly a long article. 128K is a book. 1M is ten books.
Behind this race lies a shared belief: the next breakthrough in AI is not how "smart" it is, but how much information it can handle in one go. Real-world problems are never isolated. A contract must be checked against statutes. A bug requires understanding the whole codebase. A medical diagnosis requires decades of records.
Once context windows cross the million-token mark, AI shifts from "chat tool" to "research assistant." That transformation cannot be overstated.
Open Source vs. Closed Source: Two Parallel Universes
After the refactor, opening any vendor JSON reveals an openSourceStatus field that splits the AI world in two.
Open-source camp: DeepSeek (nearly everything), Alibaba (Qwen3.5 series), Meta (Llama series), Zhipu AI (part of ChatGLM), Stability AI (image models).
Closed-source camp: OpenAI, Anthropic, Google, Baidu, Moonshot, ByteDance (Seed series).
Some companies straddle both. Alibaba is the clearest example: Qwen3.5 open-source alongside Qwen3.7 closed-source. Meta is similar; Llama is open-source, but the latest multimodal models may not be.
Behind this divide is a different business logic. The closed-source camp believes the model is the product and API calls are the revenue. The open-source camp treats the model as a customer-acquisition tool: build influence through open releases, then monetize through cloud services, enterprise editions, and technical support.
Both paths work. As of 2026, the open-source camp appears to have the louder voice. DeepSeek's R1 and Alibaba's Qwen iterate at breakneck speed. The closed-source camp still leads on raw capability (GPT-5, Claude Opus 4.8, Gemini 3.5), but that lead is shrinking.
Model Tags: A Five-Pointed Star of Capability
Every model carries a set of modelTags, like a character stat screen:
- Text generation: the baseline, present on nearly all models.
- Deep thinking: reasoning, the headline tag for DeepSeek-R1, Claude Opus, and the o-series.
- Visual understanding: image-grounded comprehension, standard on GPT-4o, Gemini, Qwen3.5.
- Code enhancement: writing and fixing code, a programmer's necessity.
- Tool calling: invoking external tools (search, calculators, databases), a key step toward agentic AI.
The Deeper Meaning of the Refactor
Back to the code change itself.
Three files became 19. Capability-based categorization became vendor-based. This is not merely an engineering optimization. It reflects an important shift in the AI industry: a model's identity is moving from "what I can do" to "where I come from."
This is no accident. In 2026, choosing a model increasingly means choosing an ecosystem.
Use an OpenAI model? You are likely using its API, Playground, and broader ecosystem. Use DeepSeek? You are probably self-hosting, because it is open source. Use Alibaba's Qwen? You may be on Alibaba Cloud.
Models are no longer isolated tools; they are gateways to ecosystems. Sorting by vendor is, in effect, an acknowledgment of that reality.
Epilogue: A Living Map
The easy-learn-ai project is, at heart, a map.
Once, that map was drawn by terrain: text region over here, image region over there, video region further on.
Now, it is drawn by country: OpenAI's territory here, DeepSeek's domain there, Alibaba and Google in the distance.
Neither drawing is wrong. But the second answers a question the first cannot: if you need to pick a model today, where do you look?
The answer is: go to your vendor.
In this great AI age of exploration, models are no longer just models. They are flags, territories, expanding empires.
And these 19 JSON files are the most honest snapshot of that map.
---
*This article is based on analysis of the easy-learn-ai project commit e6c189a.*
#easy-learn-ai #daily-update #memory #xiaokai #AI-model #open-source #large-models