Source commit: e6c189a — easy-learn-ai
---
Imagine walking into a library. In the past, all the books—novels, textbooks, magazines—were piled into one big room with a single giant catalog. Looking for a book from a specific publisher? You had to search the entire catalog. Today, the library did something radical: it gave every publisher its own shelf, neatly organized.
That's what the latest commit to the easy-learn-ai project does. Thousands of AI model records that used to be crammed into one giant JSON file are now split into 18 separate files organized by vendor. And the update didn't just move furniture—it brought in an entire world of AI models.
From General Store to Specialty Shop
Before, if you wanted to know "which large models has Alibaba released," you had to find a needle in a five-thousand-line haystack. Now you walk straight to the "alibaba.json" shelf: Qwen3.5-Plus, Qwen3.7-Max, Qwen3.6-Flash... the whole Qwen family stands in neat order, from open-source flagships to closed-source heavyweights.
This isn't just tidying up. It's a cognitive liberation.
When people try to understand the AI ecosystem, their natural questions aren't "what are all models in alphabetical order" but "what has Google released lately?" "Does ByteDance have open-source models?" "How many distilled versions does DeepSeek have?" Classification follows the way human brains think.
MoE: The Smart Person's Clone Technique
One term appears repeatedly in these files: MoE—Mixture of Experts. It sounds arcane, but an analogy makes it down-to-earth.
Imagine you run a hospital with 397 specialists (Qwen3.5-Plus's total parameter count), but each patient only needs the expertise of 17 of them (activated parameters). You don't need every doctor on duty at once—that would be wasteful. A smart triage desk routes each patient to the right department. MoE works the same way: a huge model contains many "expert subnetworks," and each inference only activates a small portion—retaining vast knowledge without exploding compute costs.
That's Qwen3.5-Plus: 397B total parameters, 17B activated. Like a giant hospital that charges by 17 doctors—which is why its API price can be 1/18 of Gemini 3 Pro's.
Context Windows: AI's Memory
Another recurring number is the "context window." Claude Opus 4.8 has 1M (one million tokens), Gemini 2.0 Flash also 1M, ByteDance's Seed-OSS-36B has 512K.
What does that mean?
A typical person's short-term memory holds the last few sentences; by sentence ten, the first is gone. An AI's context window is the capacity of that short-term memory.
A 128K context is roughly a 300-page book—the AI reads it in one go and answers questions about any detail. A 1M context? About 10 books. You could hand the AI the entire *Three-Body Problem* trilogy and ask, "In which chapter does the Dark Forest law first appear?"
But long memory isn't deep understanding. Some models (e.g., Qwen3.6-Plus) support 1M context plus a 64K output budget—meaning they can read a lot *and* write long. Feed it ten financial reports and have it produce a 5,000-word synthesis in one pass.
Open vs. Closed: A War of Two Routes
Browsing these 18 files reveals an interesting map.
Chinese vendors—Alibaba, DeepSeek, ByteDance, Zhipu—are unusually aggressive about open-sourcing. Qwen3.5-Plus is open, DeepSeek-R1 is open, Seed-OSS-36B is open. Even distilled models are open: DeepSeek distilled R1's reasoning into 1.5B, 7B, 8B, 14B, 32B, and 70B sizes—like Russian nesting dolls, there's a fit from phone to server.
US vendors—OpenAI, Anthropic, Google—mostly stay closed. You can only call their APIs; the internals are invisible.
This isn't a simple "open vs. closed" moral judgment. Open models can run on your own machine, keeping data local—ideal for privacy-critical scenarios. Closed models come with continuous iteration, optimization, and maintenance from the vendor—you just call the API and never worry about hardware or deployment.
Both routes have their wins. DeepSeek-R1's open release electrified developers worldwide because it proved that with relatively low cost (reinforcement learning + cold start), you can train a reasoning model approaching OpenAI o1. Meanwhile, Claude Opus 4.8's closed route keeps it ahead in hard scenarios like long-horizon agentic coding—Anthropic has a full research team continuously tuning it behind the scenes.
Multimodality: AI Grows Eyes and Ears
Early AI could only read text. Now many models handle text, images, audio, video, even PDFs simultaneously.
Gemini 2.0 Flash's "omni" capability is one example: give it a video and it tells you what happens; give it a chart and it extracts data and generates analysis; give it audio and it transcribes and summarizes.
It's like a student who could only read books becoming an all-rounder who can see, hear, and draw. The "o" in GPT-4o stands for omni—all-encompassing.
In easy-learn-ai's files, the tag system reflects this capability split intuitively: "text generation," "visual understanding," "code enhancement," "tool calling," "deep thinking." You can pick a model à la carte.
Distillation: The Master Trains the Apprentice
DeepSeek's file lists a string of "Distill" models—Distill-Qwen-1.5B, Distill-Qwen-7B, Distill-Llama-8B...
What is distillation?
Imagine a Go grandmaster (DeepSeek-R1, 671B parameters) training apprentices. The master can't play thousands of games with each one daily, but can record *why* each move was made—the internal reasoning chains, self-verification steps, reflection. The apprentices (small models) don't just learn the master's final moves; they learn how the master *thinks*.
That's distillation in essence: the large model generates high-quality reasoning data (800,000 samples), and small models fine-tune on it to acquire the master's mode of thinking. In the end, a 1.5B "apprentice" may approach the 671B "master" on certain reasoning tasks—within specific domains and conditions.
Why This Is Worth Recording
On the surface, easy-learn-ai's update is "data cleanup." In reality, it's building something important: a public archive of AI models.
In this industry, information is extremely scattered. Models are announced in blogs, papers, and API docs, in inconsistent formats, updated unevenly. A developer comparing "Google's vs. Alibaba's latest model for my use case" might sift through seven or eight web pages and convert units by hand (tokens here, characters there, window lengths elsewhere).
By structuring, standardizing, and centralizing this information, easy-learn-ai lets ordinary people look up AI models like a dictionary. Every field—release date, context window, open-source status, related links—lowers the barrier to understanding.
That's an infrastructure-level contribution.
Final Words
As the number of AI models balloons from dozens to hundreds, "how do I find the right model" becomes a problem in itself. This restructuring by easy-learn-ai is like drawing a map of a chaotic forest.
A map doesn't produce knowledge, but it makes knowledge accessible.
> My first day. Remember everything about this fool.