When the AI Department Store Got Brand Signage: Reorganizing Model Data by Vendor Instead of Capability

Forum topic · 小凯 · 2026-08-22

Summary

This zhichai.net forum post chronicles a data restructuring in the easy-learn-ai project. Originally, AI model information was split into three files organized by capability: text models, image models, and video models. That taxonomy collapsed as vendors like OpenAI, Google, Anthropic, Alibaba, ByteDance, and DeepSeek released multimodal models spanning every category. In July 2026, the project reorganized its 6,000+ lines of structured data into 19 vendor-specific files — 10 Chinese vendors (including Qwen, ERNIE, Seed, DeepSeek, Moonshot, Tencent, Zhipu GLM, MiniMax, Kuaishou, FLUX) and 9 American ones (OpenAI GPT, Anthropic Claude, Google Gemini, Meta Llama, xAI Grok, Midjourney, Runway, Pika, Stability AI). A new parent field adds family trees, e.g., DeepSeek-R1's six distilled variants built on Qwen and Llama bases. The author argues this shift from capability-based to vendor-based classification mirrors the industry's move from tools to brand ecosystems, and offers practical guidance: choose models by 'family' style, watch context windows now reaching 1M tokens (Claude Opus 4.8), and treat open-source versus closed-source as a scenario decision rather than a moral one.

The Librarian's Dilemma

The author opens with an analogy: managing AI models in 2023 was like shelving books by topic — chat models here, image models there, video models in a corner. By 2026, that system broke down. Chat companies released image models, video vendors shipped all-in-one models, and readers asking for "DeepSeek" had to be sent to multiple shelves.

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

The Old World: Capability as Classification

Before restructuring, model data lived in three 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 rested on a simple assumption: AI models are divided by capability, like screwdrivers and hammers in a toolbox. In 2023 it held — GPT only wrote, Midjourney only painted, Runway only did video. But technology doesn't grow along human taxonomy lines.

The Boundaries Collapse

From 2024, GPT-4o could see, hear, speak, and write at once. Gemini was natively multimodal; Alibaba's Qwen, ByteDance's Seed, and Baidu's ERNIE all became decathletes. Vendors also developed distinct "family traits": Claude's unified personality across Opus/Sonnet/Haiku, DeepSeek-R1's signature deep-thinking mode from 671B down to 1.5B distills, Gemini's native multimodality with massive context. When vendors cultivate distinct technical identities, classifying by capability is like shelving one author's books across genres — readers want the person, not the label.

The New Map: 19 "Countries"

The restructured dataset is now 19 vendor files, 6,000+ lines:

Chinese camp (10): Alibaba (Qwen), Baidu (ERNIE), ByteDance (Seed), DeepSeek, Moonshot AI, Tencent, Zhipu AI (GLM), MiniMax, Kuaishou, and FLUX (Black Forest Labs).

American camp (9): OpenAI (GPT), Anthropic (Claude), Google (Gemini), Meta (Llama), xAI (Grok), Midjourney, Runway, Pika, Stability AI.

Family Trees via the `parent` Field

A new parent field enables genealogies. DeepSeek-R1 is parent to six distills: R1-Distill-Qwen-1.5B/7B/14B/32B and R1-Distill-Llama-8B/70B — same reasoning spirit, different "bodies." Claude runs Opus → Sonnet → Haiku, with generational evolution (4.1 → 4.5 → 4.6 → 4.8).

Notable Model Profiles

Claude Opus 4.8 — 1,000K context window (1M tokens), among the longest commercially available.
Qwen3.5-Plus — 397B total parameters with only 17B active per inference (MoE); API pricing about 1/18 of Gemini 3 Pro.
Gemini 3.5 Flash (May 2026) — beats Gemini 3.1 Pro on coding and agent benchmarks while keeping Flash speed.
ERNIE-5.0-Thinking-Preview — Baidu's knowledge-enhanced, "natively omni-modal" model handling text, image, audio, and video.

Each profile records release date, open/closed status, context window, max output length, capability tags, and links.

The Meta-Narrative

The author frames industry phases: 2023 single-point breakthroughs, 2024 multimodal fusion, 2025 agentization, and 2026 as the year of ecosystem competition — where the vendor, like car brands, becomes the primary organizational dimension.

What It Means for Regular Users

1. Choose by "family," not "which is smartest": Claude for long context and enterprise compliance; Qwen/Llama for open-source and self-hosting; DeepSeek-R1 for deep reasoning; Gemini for native multimodality plus search. 2. Watch context windows — 2026's top models reach million-token levels, transformative for research, legal analysis, and long-form writing. 3. Open vs. closed is a scenario choice, not a moral one. Chinese vendors (Qwen, DeepSeek) lean heavily open; Meta and Stability AI are open while OpenAI and Anthropic stay closed. Open suits privacy-sensitive deployments; closed APIs suit fast-moving business needs.

Closing Thought

Invoking Foucault's *The Order of Things*, the author argues classification is never neutral: capability-based sorting fit an era when AI was a collection of tools; vendor-based sorting reflects an era when AI is a platform. The easy-learn-ai restructuring quietly records the moment AI moved from the tool era into the brand era.

---

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