Mapping the AI World: When Someone Decides to Draw a Complete Map of AI
> Source: easy-learn-ai project, commit e6c189a
> Date: 2026-07-12
Imagine living on a new continent that expands every day — new mountains rise, rivers shift, cities appear overnight. Your map is a single parchment where everything is crammed together: towns, mountains, rivers, forests, all overlapping, borders blurred. Want to find the city of "DeepSeek"? You have to read from start to finish. Want to know how far "Claude" is from "Gemini"? You have to decipher the chaos yourself.
This isn't a fable. It's the real situation of anyone trying to understand the world of AI models in recent years.
---
The Problem with One Parchment
In the easy-learn-ai open-source project, all AI model information was long stuffed into a single file: src/utils/model.json. Over 5,000 lines. Print it out and it would cover a small table.
What's in those 5,000 lines? OpenAI's GPT series, Anthropic's Claude family, Google's Gemini legion, China's DeepSeek, Qwen, ERNIE... all dumped into one bucket like building blocks of different colors and shapes.
Updating any model meant searching through thousands of lines. Comparing vendors meant constant scrolling and Ctrl+F. For a newcomer asking "what AI models exist?", the view wasn't a clear map — it was a parchment splashed with ink.
Not the maintainers' fault. When there were only a few dozen models, one JSON file was fine. But as the count passed one hundred, then two hundred, approaching three hundred, keeping everything in one file was like trying to fit a whole library into one suitcase.
So someone decided: it's time to draw a real map.
---
From Chaos to Order: Splitting the Continent into Territories
This commit (e6c189a) did something simple but profound: split that 5,000-line file into twenty independent files, one per vendor, under src/data/models/. Like a cartographer finally deciding not to draw every country's borders on one sheet, but to give each country its own map.
The American camp: arms race among closed-source giants
OpenAI — 38 models, the largest territory on the map.
From GPT-4 (8K context) in 2023 to GPT-5.5 (1M context) in 2026, OpenAI's model line is an ever-extending highway. GPT-4o's "omni" concept lets it understand text, audio, images, and video simultaneously. What's most interesting is the product strategy: a large/medium/small tier structure. The GPT-4.1 series (nano/mini/standard) works like a well-organized team — nano handles fast interceptions (lowest latency, lowest cost), mini runs the midfield (balanced performance and speed), and the standard version takes the final shot (highest intelligence).
Anthropic — 12 models, the boutique route.
Claude's naming itself is interesting: Opus, Sonnet, Haiku — corresponding to long, medium, and short, hinting at their positioning. Opus is the flagship with 1M context and 128K output for tasks requiring "deep thinking." Sonnet is the value pick, sometimes even surpassing Opus. Haiku is the light cavalry — fast and cheap, suited for real-time conversation.
Claude 4.8, released May 2026, introduced a special capability called "adaptive thinking": like an experienced veteran, it answers simple questions quickly and automatically slows down to reason deeply on complex ones. This "knowing when to be fast and when to be slow" is precisely what many AI models lack.
Google — 15 models, a firm bet on multimodality.
Gemini was designed around multimodality from the start. 2.0 Flash supports 1M context and handles text, images, audio, video, and PDFs. 3.1 Pro scored 77.1% on ARC-AGI-2 — a benchmark dubbed "AI's IQ test," where a high score implies near-human abstract reasoning. Gemini 3.5 Flash is even more interesting: officially it "surpasses Gemini 3.1 Pro on coding, agentic, and multimodal understanding benchmarks" while keeping the Flash series' speed — like a sprinter discovering they also win marathons.
Meta — 7 models, the lighthouse of the open-source world.
The Llama series may be the most impactful open-source model family of the past two years. From Llama 2 to Llama 4, Meta has insisted on open-sourcing, letting researchers and developers worldwide freely use, modify, and deploy these models. If the AI model world were a Cold War space race, Meta would be the one publishing the rocket blueprints.
xAI — 9 models, Musk's ace in the hole.
Grok carries a distinctive xAI flavor: real-time information access (via the X platform), long context (1M tokens), and a deliberately emphasized "humorous" personality. Grok 3.5, released April 2026, billed itself as "the smartest AI on Earth" — a very Musk form of marketing.
The Chinese camp: dual tracks of open and closed source
Alibaba (Qwen/通义千问) — 28 models, the vanguard of open source.
Qwen3.5-Plus deserves a close look: 397B total parameters but only 17B activated, thanks to the MoE (Mixture of Experts) architecture. Picture a large hospital with 397 departments where only 17 consult on any given patient — huge capability at greatly reduced running cost. Even more striking is its value: API pricing at 1/18 of Gemini 3 Pro. On the performance-price plane, Qwen3.5-Plus may be the point furthest to the upper-left.
DeepSeek — 17 models, the benchmark for open-source reasoning.
DeepSeek-R1 may have been the most sensational release of early 2025. It open-sourced not only the main model but a series of "distilled" versions from 1.5B to 70B, covering deployment from Raspberry Pi to server. "Distillation" is an elegant idea: a top Go master (R1) plays tens of thousands of games and hands the records to students of different levels. The 1.5B version is like a kindergartner learning basic openings; the 70B version like a pro grasping deep strategy. Crucially, every student gets the game records for free.
Zhipu AI (智谱) — 24 models, representative of China's LLM startups.
The GLM (General Language Model) series uses a distinctive autoregressive-blank-filling architecture, unlike GPT's pure autoregression. The latest GLM-4.5 series supports 1M context with strong coding, reasoning, and tool-calling. Zhipu's product line is broad: from ultra-light GLM-Z1-1B (runs on phones) to flagship GLM-4.5-Plus (deep thinking and visual understanding).
Baidu (ERNIE/文心一言) — 13 models, the persistence of knowledge enhancement.
ERNIE's core feature is "knowledge enhancement" — injecting encyclopedic knowledge and structured data directly into training. ERNIE-5.0 is "natively omni-modal," understanding and generating text, images, audio, and video. Baidu's strategy is clear: not the most radical, but the most practical — targeting enterprise Q&A, coding, and content creation.
ByteDance, Tencent, Moonshot, MiniMax, Kuaishou — China's second tier.
ByteDance's Seed series, Tencent's Hunyuan, Moonshot's Kimi, MiniMax's abab, Kuaishou's Kling — each with its own specialty. Kimi is known for ultra-long context (earliest with 200K Chinese context), Kling shines at video generation, and Hunyuan focuses on multimodality and enterprise applications.
Vertical specialists: image, video, audio
- Black Forest Labs — the FLUX image generation series, from a startup founded by ex-Stable Diffusion team members; a new benchmark in image quality and speed.
- Midjourney — one model, but perhaps the most famous AI name in art circles; a Discord interface turned image generation into a social experience.
- Runway — a video generation pioneer; Gen-3 moved text-to-video from lab to practice.
- Pika — another video generation company, known for speed and ease of use.
- Stability AI — creator of Stable Diffusion, synonymous with open-source image generation.
- GPT-4 (2023): 8K context
- GPT-4o (2024): 128K context
- GPT-4.1 (2025): 1M context
- Closed-source camp: OpenAI, Anthropic, Google, Baidu (ERNIE), ByteDance (Seed), xAI — models as black boxes you can use but can't inspect.
- Open-source camp: DeepSeek, Alibaba (Qwen), Meta (Llama), Zhipu (GLM) — weights, architecture, and training methods published; anyone can download, modify, and deploy.
- 2023: GPT-4 bursts onto the scene — the first year of large models.
- 2024: Multimodality explodes, led by GPT-4o and Gemini 1.5.
- 2025: Reasoning models rise — DeepSeek-R1 and Claude 3.5 Sonnet redefine "thinking."
- 2026: Context windows break a million tokens; Agents become the keyword; models start to "act," not just "talk."
---
Four Big Trends Behind the Data
Put 200+ models side by side and clear patterns emerge.
Trend 1: Context windows are exploding
Three years, a 125-fold increase. "Context window" is a technical term for how much an AI can "remember." 8K is a short story; 128K, a novel; 1M, the entire Harry Potter series. It means AI is evolving from "recite a passage and answer" to "read your company's entire document set, then answer." A 1M-context model can swallow thousands of pages of legal contracts, tens of thousands of lines of code, or a mid-sized enterprise's whole knowledge base.
Trend 2: Multimodality becomes standard
Early models could only process text — like someone who could read but not see. Today's flagships — GPT-4o, Claude Opus 4.8, Gemini 3.5 Flash, Qwen3.7-Max — all support visual understanding: images, charts, video frames, some even audio in and out. This isn't garnish; it's a fundamental shift. The human world was never pure text, and AI is evolving the same way.
Trend 3: Reasoning becomes the new battleground
Before 2024, competition was about knowledge — who knows more. After 2025, it's about reasoning — who thinks better. DeepSeek-R1, Claude Opus 4.8, ERNIE-X1, the GPT-o series share one trait: they "think," deriving step by step like solving a math problem rather than reciting training data. DeepSeek-R1's chain-of-thought brought it near OpenAI o1 level on math and code. The exam changed from open-book to closed-book.
Trend 4: The open/closed divide
The dataset includes an openSourceStatus field, and reading the map reveals a clear line:
This divide isn't random; it reflects strategy. Closed vendors treat models as core competitive assets and revenue; open vendors treat them as ecosystem foundations — attracting developers with free access, locking in users through the ecosystem. And open-source models are closing the gap at astonishing speed: DeepSeek-R1 approaches or exceeds OpenAI's o1 series on multiple benchmarks; Qwen3.5-Plus outperforms trillion-parameter closed models. Unthinkable two or three years ago.
---
Why Draw This Map?
easy-learn-ai isn't a big company's official product — it's an open-source project. Why does the maintainer spend time collecting, organizing, and verifying 230+ models? The answer hides in a detail: every model entry has a relatedLinks field with three links — official announcement, technical docs, GitHub repo. This isn't data copying; it's curation.
In an age of information explosion, "finding the right information" matters more than "having information." easy-learn-ai reduces search costs: no visiting twenty websites, no bouncing between Discord, Twitter, and official blogs. Want to understand a model? Open the corresponding JSON file; everything is there.
And this refactor makes the curation sustainable. New model? Add an entry in the vendor's file. Model updated? Edit its description. No more needle-in-haystack searches across 5,000 lines. It's open-source spirit in action: someone does hard one-time work so everyone benefits.
---
What This Map Tells Us
Sort all models by release date and a clear timeline appears:
---
Final Words
A philosopher said: "The map is not the territory." However precise, a map is an abstraction of reality, not reality itself.
This model map is the same. It records parameters, context lengths, release dates — but not the "feel" of actually using the models. Claude's honesty, GPT's fluency, DeepSeek's depth — subjective, experiential qualities no JSON file can capture.
Still, a good map has value. It tells us where we are, what's around us, and what may lie ahead. On a continent that changes daily, having a map is better than not having one.
---
*This article is based on data from easy-learn-ai commit e6c189a. The project catalogs 230+ AI models from 20 vendors, spanning text generation, visual understanding, code enhancement, tool calling, and more.*