Tencent Hunyuan glossary
No procedures. If you only need a door, use the learning map.
Hy / Hunyuan
What it is: Tencent’s current brand for the self-developed model family. hunyuan.tencent.com and hy.tencent.com are the same site. International copy has said “Tencent Hy, formerly known as Tencent Hunyuan”.
What it is not: Yuanbao the app, a TokenHub account, or an executable.
Repos, papers, and the older cloud product still say Hunyuan. Both names must map to an official page. Do not invent a third product name like “Hunyuan 3.0” — the current flagship LM is Hy3.
Internals (why MoE, what MTP is) belong in Learn LLM.
Hy3
What it is: Flagship language model shipped 2026-07-06. TokenHub slug hy3. Open weights at Tencent-Hunyuan/Hy3: 295B total, 21B active, 256K context, Apache-2.0.
Hosted capabilities (model list): retained thinking, structured output, function calling, cache. The product page adds Coding / long-context / Agent.
What it is not:
- Not
hy3-preview(official retire date 2026-08-31). - Not the image / video / 3D models.
- Not a button inside CodeBuddy.
TokenHub
What it is: Tencent Cloud’s “LLM service platform”, product tree 1823. Official overview: one door that bundles Hunyuan and third-party models. Calling Hy is one line of business.
Why the name matters: In 2026 the Hy3 call guide, prices, keys, and Claude Code wiring all live on this tree. Calling it “the Hunyuan console” hides the multi-vendor catalog and collides with tree 1729.
Auth is a TokenHub API key, not a random CAM SecretId/SecretKey.
Two cloud doc trees
| Tree | ID | How to use it now |
|---|---|---|
| TokenHub | 1823 | Primary source here. Hy3, prices, OpenAI-compatible endpoint |
| Tencent Hunyuan LLM | 1729 | Still online. Still shows older Hunyuan-T1 / TurboS specs |
cloud.tencent.com/product/hunyuan still scrapes as the older lineup; product/tclm already features Hy3. When marketing pages disagree, TokenHub docs + hunyuan.tencent.com win, and the conflict is named in the prose.
Why thinking has three vocabularies
Not three products — three official pages, three JSON shapes:
| Source | How it is spelled |
|---|---|
| TokenHub call guide | thinking: { type: "enabled" }; reasoning_effort (default low on hy3) |
| tclm product page | no_think / think_low / think_high |
| Hy3 open README | extra_body.chat_template_kwargs.reasoning_effort = no_think | low | high (default no_think) |
This site does not guess that TokenHub low equals a particular open-weight enum. Copy the fields from the page you are actually calling. With tools present, TokenHub maps low to high (call guide).
Sibling products are not Hunyuan
| Name | Official positioning | Entry | This site |
|---|---|---|---|
| Yuanbao | “Tencent’s all-in-one AI assistant… already using the latest Hunyuan Hy3 model” | yuanbao.tencent.com | #76 |
| CodeBuddy | Tencent Cloud coding assistant (IDE / CLI) | codebuddy.cn | #78 |
| WorkBuddy | Office agent workbench | workbuddy.cn | Map row |
| ima | “AI knowledge steward based on a knowledge base” | ima.qq.com | Map row |
They consume Hunyuan (or other models). They are not “Hunyuan open capabilities”. The Hy3 README names CodeBuddy / Cline / KiloCode as evaluation scaffolds. That does not make Hunyuan equal to CodeBuddy.
Open-source names vs API slugs
The Tencent-Hunyuan org still uses Hunyuan* repo names (HunyuanVideo, HunyuanImage-3.0, Hunyuan3D-2.1). TokenHub call parameters prefer hy-* / hy3.
Reconcile against the page you are reading. Do not treat similar names as the same model field.
Internals this handbook will not teach
MoE routing, MTP speculative decoding, quantization training, RL post-training: the README has sections; they are model internals. Use LLM fundamentals. This handbook only keeps official spec numbers and serve commands.