When the AI World Went From One Encyclopedia to Twenty Libraries: A Per-Vendor Model Data Refactor

Forum topic · 小凯 · 2026-08-18

Summary

This post describes a refactor of the easy-learn-ai project: a single 5,000+ line file cataloging AI models was split into 20 vendor-specific files, a change the author frames as a mirror of the AI industry's transformation. The piece surveys the resulting dataset: DeepSeek's progression from LLM and V2/V3 MoE models to R1 reasoning models, distilled variants, and V4-Pro (1.6T total/49B active parameters, 1M context); Anthropic's Haiku/Sonnet/Opus tiering and rapid Claude 4.x cadence; OpenAI's move from GPT-4o and o1/o3 to the GPT-5 family and the Apache 2.0 licensed GPT OSS models; Google's natively multimodal Gemini line plus Imagen 4, Nano Banana, and Veo 3 media generation; Alibaba's aggressive Qwen3/Qwen3.5 open-source matrix; and ByteDance's Seed and Doubao families. The author distills five industry patterns: model-building is no longer a three-player game, MoE architecture has become mainstream, context windows are in an arms race (approaching 1M tokens), reasoning-trained models are now consensus, and open-source models are closing the gap with closed ones.

Key points

The author describes a single commit in the easy-learn-ai project that split one 5,000+ line file containing AI model data into 20 vendor-specific files. Beyond code hygiene, the change is read as an honest mapping of an industry that has grown from a handful of models to 20 vendors spanning text, image, video, audio, and code.

The twenty "libraries" — notable shelves

DeepSeek: a clear iteration chain — DeepSeek LLM (Jan 2024, 7B/67B), V2 (May 2024, MoE with 236B total / 21B active), V3 (Dec 2024, 671B total / 37B active), R1 (Jan 2025, first reasoning model trained with RL), V3.1/V3.2-Exp (dual thinking/non-thinking modes), and V4-Pro (Apr 2026, 1.6T total / 49B active, 1M context). Its R1-Distill-Qwen (1.5B–32B) and R1-Distill-Llama (8B/70B) models transfer reasoning via 800K reasoning samples. DeepSeek-Math-V2 reportedly reached gold-medal level at IMO 2025 and scored 118/120 on Putnam 2024.
Anthropic: disciplined product tiers — Haiku (fast/cheap), Sonnet (balanced), Opus (flagship) — with a regular cadence: 3.5 series (Jun–Oct 2024), 3.7 Sonnet (first hybrid reasoning model, Feb 2025), the 4 series (May 2025), and 4.1–4.8 upgrades every two-to-three months. Claude Sonnet 4.6 reportedly topped Humanity's Last Exam and led GPT-5.2 by ~144 Elo on GDPval-AA.
OpenAI: from GPT-4 and GPT-4o (omni multimodal) through the o1/o3 reasoning line to the unified GPT-5 family (mini/nano/pro/codex) and GPT-5.5 (1.05M context, 128K output). A landmark: GPT OSS, open-sourced under Apache 2.0 (120B and 20B MoE, 4-bit quantization, compatible with Transformers, vLLM, llama.cpp, Ollama).
Google: natively multimodal Gemini (2.0 Flash → 3.5 Flash, ~1M token context) plus a media generation matrix: Imagen 4 (2K, 10x speed mode), the Nano Banana image-editing series (up to 4K), and Veo 3/3.1 video with native audio.
Alibaba (Qwen): the most aggressive open-source matrix — Qwen3-235B-A22B (thinking/non-thinking), Qwen3-30B-A3B, Qwen3-4B, Qwen3-VL, Qwen3-Coder-480B-A35B, QwQ-32B — plus Qwen3.5-Plus (397B total / 17B active, open, API priced at ~1/18 of Gemini 3 Pro), and Wan2.2 video, Qwen-Image, and Z-Image (6B, 8-step generation).
ByteDance: Seed-OSS-36B (512K context), the doubao-seed 1.6/2.0 families (pro/lite/flash/thinking/code), Seedream text-to-image (up to 4K), Seedance video (multi-shot narrative), and doubao-seed-1.6-thinking with a 32K chain of thought.
Newer names to watch: Black Forest Labs (FLUX), xAI (Grok), Moonshot/Kimi (2M context), MiniMax, Runway/Pika.

Five industry patterns the data reveals

1. No longer a three-player game — 20 vendors, with Chinese companies (Alibaba, ByteDance, Baidu, DeepSeek, Moonshot, MiniMax, Zhipu, Tencent, Kuaishou) accounting for roughly half. 2. MoE is the mainstream architecture — DeepSeek-V4 (1.6T/49B), Qwen3.5-Plus (397B/17B), and GPT OSS all use sparse expert activation to balance capability and inference cost. 3. Context-window arms race — 64K → 128K → 256K → 1M → 1.05M (GPT-5.5: 1,050K; Gemini 3.5 Flash: 1,048,576; DeepSeek-V4-Pro: 1M), pushing AI from short-term toward long-term memory. 4. Reasoning goes mainstream — from DeepSeek-R1 and OpenAI o1/o3 to Claude's hybrid reasoning and Qwen thinking modes: training has shifted from next-token prediction toward RL-trained chains of thought. 5. Open vs. closed convergence — DeepSeek, Qwen, Llama, Gemma, and now GPT OSS vs. GPT-5.5, Claude, and Gemini; DeepSeek-V4-Pro is positioned as a near-frontier open model.

Conclusion

The refactor is framed as more than paying down technical debt: when one file becomes twenty, the data structure itself becomes information — a living map of a multipolar, layered, ecosystem-like AI world, updated daily.

Source

easy-learn-ai project commit; vendor official documentation and model cards.