Key points
A recent commit to the easy-learn-ai project (commit e6c189a) replaced one monolithic 5,000-line model.json with 20 per-brand JSON files under src/data/models/, turning a cluttered "grocery store" of AI model data into a navigable map. Each file covers one vendor's lineup — parameters, context windows, capability tags, release dates, and links.
The 20 brand files
alibaba.json, anthropic.json, baidu.json, black-forest-labs.json, bytedance.json, deepseek.json, google.json, kuaishou.json, meta.json, midjourney.json, minimax.json, moonshot.json, openai.json, pika.json, runway.json, stability-ai.json, tencent.json, xai.json, zhipu-ai.json, and more — each independently maintainable, so new models can be added without breaking unrelated vendor data.
Highlights from the tour
- DeepSeek: R1-Distill series (1.5B–70B) for local inference with full chain-of-thought skeletons; V4 Pro/Flash use a 1.6T-parameter MoE activating only 49B per pass.
- Moonshot (Kimi): from moonshot-v1-8k to K2 (1T-parameter MoE), K2.5 (Agent Swarm with ~100 sub-agents), and K2.6 (multimodal agent). Kimi-Linear-48B compresses KV cache by 75% via linear attention for 6x faster decoding.
- OpenAI: density-tiered lineup from GPT-3.5 through GPT-5.5 (1050K context), GPT-5.5 pro, GPT-5.3-Codex, and surprisingly open GPT OSS (120B/20B, Apache 2.0). GPT-5 introduces adjustable "minimal reasoning" effort levels.
- Alibaba (Qwen): Qwen3.5-Plus — 397B total / 17B active — outperforms the trillion-parameter Qwen3-Max at roughly 1/18 the API price of Gemini 3 Pro, with full size coverage from 0.5B to 3.7-Max.
- ByteDance (Seed): open-sourced Seed-OSS-36B-Base with 512K context and 12T training tokens plus adjustable thinking budget; Doubao-Seed-Code targets developers with 256K context, image/video understanding, and tool calling.
- easy-learn-ai project commit e6c189a
- DeepSeek API docs: https://api-docs.deepseek.com
- Kimi platform: https://platform.moonshot.cn
- OpenAI models docs: https://platform.openai.com/docs/models
- Gemini API docs: https://ai.google.dev/gemini-api/docs/models
- Qwen model list: https://help.aliyun.com/zh/model-studio/text-generation-model/
Three trends the map reveals
1. 1M-token context is the new baseline — DeepSeek V4, Gemini 2.0 Flash, Kimi-Linear-48B, GPT-4.1/5, and Qwen3.6-Plus all reach ~1M tokens (about 750,000 Chinese characters, or three copies of *Dream of the Red Chamber*).
2. MoE is standard equipment — DeepSeek V3 (671B/37B active), Kimi K2 (1T/32B), Qwen3.5-Plus (397B/17B): hundreds of specialists with a router, like a hospital's department-based referrals.
3. Thinking mode is a standard dial — DeepSeek V3.2/V4 Thinking/Non-Thinking, OpenAI reasoning.effort from none to xhigh, Kimi K2-thinking, ERNIE-5.x-Thinking: users can now tune depth of reasoning per task.
Why it matters
The refactor is less about code hygiene than about making the exploding model ecosystem legible: developers can compare and choose models at a glance, and the modular structure stays sustainable as new vendors and models ship weekly. As the post puts it: *"The map is not the territory — but without a map, you don't even know where you are."*