Tencent Hy4 preview (TokenHub)
Hy4 preview is Tencent Hy Team's
next-generation language model (770B MoE, 49B active, 1M context). OpenTryOn
exposes it on understand — text in, text out. It is not a virtual
try-on or video-generation model.
Hosted path: Tencent Cloud TokenHub
OpenAI Chat Completions (hy4-preview). Local / open-weight serving:
Hy4 local (vLLM / SGLang).
| Registry id | hy4-preview |
| Adapter | Hy4Adapter |
| Auth | TOKENHUB_API_KEY (alias TENCENT_TOKENHUB_API_KEY) |
| Default base URL | https://tokenhub-intl.tencentcloudmaas.com/v1 |
| License (weights) | Apache-2.0 on Hugging Face |
Do not use OpenRouter or other third-party hosters for this adapter — first-party TokenHub only.
CLI
opentryon understand --model hy4-preview \
--prompt "Write a 3-sentence lookbook caption for a linen trench." --dry-run
opentryon understand --model hy4-preview \
--prompt "Describe this outfit." --image garment.jpg --reasoning-effort high
--image is optional (OpenAI vision part). Hy4's public API is a text LLM;
if TokenHub rejects the image part, omit --image. There is no --video.
--no-thinking sets TokenHub thinking.type=disabled. Default reasoning depth
is high (temperature=0.9, top_p=1.0).
Python
from tryon.api import Hy4Adapter
adapter = Hy4Adapter() # TOKENHUB_API_KEY
result = adapter.understand(prompt="Summarize this merchandising brief.")
print(result["text"])
print(result.get("reasoning"))
MCP
Tool: understand_hy4_preview. Restart MCP so TryOn Studio lists it. Connect
saves TOKENHUB_API_KEY.