When AI's Department Store Gets Brand Signboards: Reorganizing Model Data by Vendor

Forum topic · 小凯 · 2026-08-22

Summary

A detailed account of how the easy-learn-ai project restructured its AI model dataset in July 2026, moving from capability-based classification (text/image/video JSON files) to vendor-based organization with 19 provider files covering 6,000+ lines of structured data. The post argues that as models became multimodal and developers cultivated distinct 'technical personalities'—Claude's safety alignment, DeepSeek-R1's deep-thinking mode, Gemini's native multimodality—vendor identity became a more meaningful organizational principle than capability. It highlights new hierarchical data features like parent-child relationships for distilled models (DeepSeek-R1 and its six distill variants), profiles notable models such as Claude Opus 4.8's 1M-token context window and Qwen3.5-Plus's cost-efficient MoE architecture, and frames the shift as AI entering a brand-and-ecosystem competition era. Practical guidance covers choosing model families, watching context window sizes, and weighing open-source versus closed-source tradeoffs.

When AI's Department Store Gets Brand Signboards

1. A Librarian's Confusion

Imagine you are a librarian. In 2023, you received your first batch of "AI model" books. There weren't many, so you sorted them by topic: one shelf for "chat-capable" models, another for "image generation," a corner for "video." Simple, intuitive, sufficient.

A year later, problems emerged.

The chat company suddenly released an image model. A vendor once filed under video could now write articles, draw, and edit video simultaneously. Stranger still, some "books" defied any category—they did a bit of everything.

Your library descended into chaos. A reader walked in asking for "that hot DeepSeek thing," and you had to explain: DeepSeek-R1 sits in the text section, but DeepSeek's image model is in the image section... The reader's expression shifted from confusion to sympathy.

This was the real dilemma facing the easy-learn-ai project in July 2026.

2. The Old World Order: Capability as Classification

Before the refactor, the project's model data lived in three large files:

model.json — 5,000+ lines of "text models"
img.json — 600+ lines of "image models"
video.json — nearly 500 lines of "video models"

This scheme rested on a simple assumption: AI models are divided by capability, like screwdrivers, wrenches, and hammers in a toolbox.

In 2023, this assumption largely held:

OpenAI's GPT series only output text
Midjourney only drew pictures
Runway focused on video

Each stayed in its lane. Classification by capability was clear, reasonable, intuitive.

But technology never grows according to human taxonomies.

3. Borders Collapse: Every Vendor Wants to Be an All-Rounder

Starting in 2024, things changed.

OpenAI released GPT-4o, a model that could see, hear, speak, and write at once. Google's Gemini was multimodal from day one. Alibaba's Qwen, ByteDance's Seed, Baidu's ERNIE—every company transformed itself from a "single-event champion" into a "decathlete."

More subtly, vendors developed distinct "family traits."

Anthropic's Claude series—from flagship Opus to lightweight Haiku—shares the same DNA: extremely long context, cautious safety alignment, elegant chain-of-thought presentation.

DeepSeek's R1 series, from the full 671B-parameter model to distilled 1.5B variants, inherits the signature "deep thinking" mode—pausing silently to reason, then delivering self-verified answers. It's a brand temperament, not a feature label.

Google's Gemini takes another path: native multimodality, million-token context, deep search integration. Using Gemini feels entirely different from using Claude—not a difference of capability, but of worldview.

When every vendor builds its own "technical personality," classifying by capability is like shelving one author's books across different genre sections—readers seek the person, not the label.

4. The New World Map: 19 "Nations" of Models

After the refactor, data was reorganized into 19 independent files, one per vendor:

China contingent (10):

Alibaba (Qwen series) — an Eastern force in open source
Baidu (ERNIE series) — knowledge-enhanced, Chinese-native
ByteDance (Seed series) — the short-video empire's AI ambition
DeepSeek — the dark horse of reasoning
Moonshot AI — long-context rising star
Tencent — a social giant's model portfolio
Zhipu AI (GLM series) — Tsinghua-lineage technologists
MiniMax — voice and multimodal explorer
Kuaishou — vertical depth in short-video scenarios
Black Forest Labs (FLUX) — Western-headquartered but with deep Chinese-community roots among founders

US camp (9):

OpenAI (GPT series) — the pioneer that defined the track
Anthropic (Claude series) — evangelists of safety alignment
Google (Gemini series) — the search empire's AI continuation
Meta (Llama series) — heavyweight open-source player
xAI (Grok series) — Musk's speed and rebellion
Midjourney — aesthetic benchmark of art generation
Runway — video generation pioneer
Pika — rising video challenger
Stability AI — open-source image evangelist

19 files, 6,000+ lines of data. Each file is a complete dossier of one "technological nation."

5. "Micro-Politics" in the Data Structure

This reclassification reflects a deeper shift in how we understand AI.

From functionalism to brand-ism. The old scheme assumed AI is a tool, sorted by function. The new scheme assumes AI is a service, sorted by provider.

It resembles how we organize phone apps. Early on, we sorted by function: social, tools, games. Later we realized WeChat is social, payment, and mini-programs at once; Douyin is entertainment, e-commerce, and search. We started organizing by usage frequency regardless of type. Users increasingly care less about what a model "can do" and more about "whose model it is"—because vendor brands now carry enough information density.

From flat lists to family trees. The new structure supports a parent field. For example, DeepSeek-R1 is the "parent" of six "children":

DeepSeek-R1-Distill-Qwen-1.5B (distilled onto Qwen)
DeepSeek-R1-Distill-Qwen-7B
DeepSeek-R1-Distill-Llama-8B (distilled onto Llama)
DeepSeek-R1-Distill-Qwen-14B
DeepSeek-R1-Distill-Qwen-32B
DeepSeek-R1-Distill-Llama-70B

Like a genealogy: these small models are R1's "spiritual descendants," inheriting its reasoning mode with different "bodies" (parameter scale, base architecture).

Claude has a clear family tree too: Opus (flagship) → Sonnet (mainstay) → Haiku (lightweight), with generational evolution (4.1 → 4.5 → 4.6 → 4.8). This "family sense" is impossible to express with capability-based classification.

6. Technical Details Worth Savoring

Claude Opus 4.8 — a 1,000K (one million token) context window. Feed it an entire novel and it can precisely cite chapter and page. Currently among the longest commercial context windows.

Qwen3.5-Plus — 397 billion total parameters but only 17 billion activated per inference (MoE architecture), like an expert system with a vast knowledge base that only retrieves relevant parts. Remarkably cost-efficient: API pricing is 1/18 of Gemini 3 Pro.

Gemini 3.5 Flash — released May 2026, it beat Google's own Gemini 3.1 Pro on several coding and agentic benchmarks while retaining Flash-series speed, breaking the "fast = worse" stereotype.

ERNIE-5.0-Thinking-Preview — Baidu built "knowledge enhancement" into its DNA. Unlike data-driven peers, ERNIE emphasizes structured knowledge injection; version 5.0 claims "native full-modality"—understanding and generating text, image, audio, and video in one model.

Each model dossier includes release date, open/closed status, context window, max output length, capability tags, and links. Together they form a living map of the AI industry.

7. The Meta-Narrative Behind the Refactor

Zoom out, and this data restructuring mirrors the AI industry entering its "ecosystem competition" phase:

2023: "single-point breakthrough" year — who makes the best chatbot?
2024: "multimodal fusion" year — who does text, image, and video together?
2025: "Agent-ification" year — can models autonomously use tools and execute complex tasks?
2026: "ecosystem loop" year — comparing entire technology stacks, brand recognition, and developer stickiness rather than single capabilities.

Once the industry competes on ecosystems, "vendor" naturally becomes the core organizing dimension over "capability."

It mirrors the auto industry: early classification was by body type (sedan, SUV, sports car). Once every brand covered all types, buyers chose by brand—Toyota's reliability, BMW's handling, Tesla's tech. Brands carry technology路线, design philosophy, user communities—far beyond "function."

AI models are walking the same road.

8. What It Means for Ordinary Users

First: choose by "family." Don't ask "which model is strongest"; ask "what working style do I need?"

Ultra-long context, high safety, enterprise compliance? → Claude family
Open source, local deployment, cost efficiency? → Qwen or Llama family
Deep reasoning, math, and code? → DeepSeek-R1 family
Native multimodality with deep search integration? → Gemini family

Second: watch the context window. Many benchmarks compare only "intelligence," but in daily use, "how much context is remembered" often matters more. Top 2026 models have pushed context to the million-token level—enough to read dozens of papers or an entire book at once, revolutionary for research, legal analysis, and long-form writing.

Third: open vs. closed source is a scenario choice, not a moral one. Chinese vendors are notably active in open source (Qwen, DeepSeek); the US is split—Meta and Stability AI embrace openness, OpenAI and Anthropic stay closed. Open models enable private deployment and custom fine-tuning for privacy-sensitive scenarios; closed models typically update faster with more stable APIs for fast-moving business needs.

9. Closing Thought: Taxonomy Is Worldview

Foucault's *The Order of Things* explores how humans construct understanding through classification. Taxonomies are never neutral—they reflect the cognitive paradigm of their era.

In 2023 we classified AI models by "capability" because AI was still a set of tools, each with its job. In 2026 we classify by "vendor" because AI has become a platform—each vendor builds its own technological universe, and models are merely the entrance.

In a few years we may need new taxonomies—perhaps by "degree of autonomy" (assistant, collaborator, agent) or "personality traits" (cautious, creative, efficient).

Either way, this seemingly dry data refactor records a historical milestone: AI has formally moved from the tool era into the brand era.

---

*Reference: easy-learn-ai project commit e6c189a, adding 19 vendor model data files with 6,000+ lines of structured data covering OpenAI, Anthropic, Google, Meta, DeepSeek, Alibaba, Baidu, ByteDance, and other major AI vendors.*