Skip to main content

Hy4 preview (local weights)

Open weights: tencent/Hy4-preview and tencent/Hy4-preview-FP8 (Apache-2.0). Official serving is vLLM or SGLang, then an OpenAI-compatible client — the same Hy4Adapter as TokenHub, with --endpoint local.

OpenTryOn does not load 770B parameters in-process. Do not expect pip install opentryon[local] + a single GPU to run this model.

Registry idhy4-preview-local
Served model namehy4-preview
Default URLHY4_BASE_URL or http://127.0.0.1:8000/v1
VRAMFP8 checkpoint ~770GB+ weights; official recipe --tensor-parallel-size 8 (tested class: 8×B300 / 16×B200). BF16 is ~1.5TB.

Hosted twin: Hy4 TokenHub.

Serve (official)​

vLLM (FP8, 8-way TP):

docker run --gpus all \
  -p 8000:8000 \
  --ipc=host \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:hy4-preview tencent/Hy4-preview-FP8 \
    --tensor-parallel-size 8 \
    --speculative-config '{"num_speculative_tokens":3,"method":"mtp"}' \
    --attention-backend FLASHMLA_SPARSE \
    --tool-call-parser hy_v4 \
    --reasoning-parser hy_v4 \
    --enable-auto-tool-choice \
    --port 8000 \
    --served-model-name hy4-preview

See the Hugging Face card for SGLang (lmsysorg/sglang:hy4-preview).

CLI​

# HY4_BASE_URL=http://127.0.0.1:8000/v1
opentryon understand --model hy4-preview-local \
  --prompt "Hello! Briefly introduce yourself." --dry-run

Local thinking off uses the vendor chat-template flag reasoning_effort=no_think (--no-thinking).

Python​

from tryon.api import Hy4Adapter

adapter = Hy4Adapter(endpoint="local")
result = adapter.understand(prompt="Hello")
print(result["text"])

MCP​

Tool: understand_hy4_preview_local. No TokenHub key. Point HY4_BASE_URL at the vLLM/SGLang server, then restart MCP so Studio lists the model.

See also​