Tencent Hunyuan learning map
Hunyuan / Tencent Hy is Tencent's self-developed model family and the channels that open those models to developers. The research site is hunyuan.tencent.com. Hosted inference goes through TokenHub.
This directory covers models and open capabilities only. Yuanbao, CodeBuddy, WorkBuddy, and ima are sibling products. They get one row on the family map.
Product landscape
Several Tencent doors share the words Hunyuan or Hy. They are not one product with four skins.
Tencent AI (Hunyuan-related official surfaces)
├── Hunyuan / Hy — models, research, open weights (this handbook)
│ ├── Hy models (Hy3, Hy-MT2, image / video / 3D / vision / ASR)
│ ├── Research (Hy3 note, Hyra, …)
│ ├── Co-design / International (site chrome, not a call surface)
│ ├── Try Hy (Hy AI Studio)
│ └── Open weights (GitHub Tencent-Hunyuan / Hugging Face tencent/Hy3)
├── TokenHub — cloud model gateway (serves Hy and third-party models)
├── Tencent Hunyuan LLM (cloud product 1729, older API tree)
├── Yuanbao — consumer assistant (runs Hy3) → issue #76
├── CodeBuddy — coding IDE / CLI → issue #78
├── WorkBuddy — office agent
└── ima — knowledge steward| Surface | What it is | Official URL | This site |
|---|---|---|---|
| Hy models / Hunyuan site | Model and research home | hunyuan.tencent.com | This page + tutorial |
| TokenHub | Unified LLM API (Hy + third parties) | Product / docs 1823 | Tutorial / Cookbook / Cheatsheet |
| Try Hy (AI Studio) | Browser playground | aistudio.tencent.com | One tutorial section |
| Open weights | Apache-2.0 checkpoints | Tencent-Hunyuan / tencent/Hy3 | Tutorial section + cheatsheet |
| Research | Tech notes | research / Hy3 | Map row |
| Co-design / International | First-level site nav | hunyuan.tencent.com header | Map row; not a new product |
| Hunyuan cloud product (1729) | Older dedicated API tree | product/hunyuan / docs 1729 | Map row. Hy3 prose lives on TokenHub |
| Yuanbao | All-in-one assistant that already uses Hy3 | yuanbao.tencent.com | Map row → #76 |
| CodeBuddy | Tencent Cloud coding assistant | codebuddy.cn | Map row → #78 |
| WorkBuddy | Office agent workbench | workbuddy.cn / cloud product | Map row |
| ima | Knowledge-base steward | ima.qq.com | Map row |
| Jobs / face ID / CVM / SMS | Non-topic nav items | — | Out of scope |
Name collisions (full write-up in the glossary):
- Hunyuan / Hy ≠ Yuanbao. The model family is not the consumer chat app.
- Hy3 ≠ TokenHub ≠ AI Studio. Model, gateway, playground.
- There is no first-party
hunyuanCLI. The coding agent product is CodeBuddy. - Open-source repo names ≠ TokenHub
modelslugs.
Quick decision
What do I want?
├── Try Hy3 / translation / multimodal in a browser
│ └── → Hy AI Studio (aistudio.tencent.com)
├── Call Hunyuan from my own front end / Node service
│ └── → TokenHub, model = hy3 (this handbook)
├── Point Claude Code / Cursor at Hy3
│ └── → TokenHub “connect AI tools”; Hy3 walkthrough is official 1823/131903
│ └── CodeBuddy itself is #78, not this directory
├── Serve open weights
│ └── → github.com/Tencent-Hunyuan/Hy3 (official: 8-GPU class)
├── Chat / search / write, no code
│ └── → Yuanbao (#76)
├── IDE completion / repo agent
│ └── → CodeBuddy (#78)
└── CVM / COS / SMS
└── → Out of scopeSources: hunyuan.tencent.com, TokenHub docs, Hy call guide.
When this handbook is worth it
Read it when
- You want an OpenAI-compatible domestic model from TypeScript, and the official door is TokenHub
hy3. - You already use Claude Code / Cursor and want Hy3 on the same harness (official connect pages exist).
- You need to compare hosted
hy3with the open checkpoint.
Skip it when
- You want Yuanbao the chat app → #76.
- You want CodeBuddy the coding product → #78.
- You want MoE / MTP / attention internals → LLM fundamentals and Learn LLM.
Learning path
| Stage | Read | Goal |
|---|---|---|
| 1. Pick a door | This map | Stop mixing Yuanbao / TokenHub / Hy3 |
| 2. First call | Tutorial | A hy3 completion in 15 minutes |
| 3. Wire it in | Cookbook | Streaming, thinking, tools, Claude Code |
| 4. Look up | Cheatsheet | Slugs, prices, endpoints |
| 5. Names | Glossary | Which official tree wins |
Model snapshot (Hy-family slugs only)
Canonical table: TokenHub model list. Prices: pricing (2026-08-14).
| Official name | model | Window (in / out) | Official one-liner |
|---|---|---|---|
| Hy3 | hy3 | 192k / 128k (256k context) | Real-world Coding / long-context / Agent |
| Hy3 preview | hy3-preview | same | Retires 2026-08-31 |
| Hy-MT2-Pro / Plus / Lite | hy-mt2-* | 4k / 4k | Translation |
| Hy-Role / Latest | hy-role / hunyuan-role-latest | 28k / 4k | Role-play |
| HY-Image-V3.0 | hy-image-v3.0 | — | T2I / I2I |
| HY-3D-3.1 | hy-3d-3.1 | — | T23D / I23D |
Hosted Hy3: 1 / 4 / 0.25 CNY per million tokens (input / output / cache hit). Open weights are a separate path: tutorial.
Out of scope
- Full Yuanbao / CodeBuddy / WorkBuddy / ima tutorials.
- Picking DeepSeek / Kimi / GLM on TokenHub.
- Non-AI Tencent Cloud (CVM, COS, SMS, domains).
- Model internals.