🔥 Updates

• [2026-07-15] 🚀 We have released Hy-Embodied-VLM-1.0! An efficient Mixture-of-Experts vision–language foundation model for embodied agents in the physical world, activating only ~3B parameters per token (~30B total) for high inference efficiency. Weights are available on Hugging Face, together with inference code for both HuggingFace transformers and vLLM.
• [2026-06-15] 🤖 We have released HY-VLA-0.5! The official code, UMI-trained weights and 2000+ hours of high-fidelity UMI data are now available.
• [2026-04-09] 🚀 We have released HY-Embodied-0.5, featuring the open-sourced HY-Embodied-0.5 MoT-2B weights on Hugging Face along with the official inference code!

📖 Abstract

Building capable embodied agents requires not only multimodal perception and understanding, but also agentic capabilities for reasoning about actions, adapting to evolving situations, and interacting with the physical world. In this report, we introduce Hy-Embodied-VLM-1.0, an efficient and powerful embodied foundation model specifically designed for embodied agents operating in the physical world.

To cultivate such capabilities from the pre-training stage onward, we define an action-centric capability taxonomy comprising three progressive dimensions: Action-Relevant State Understanding, Action–Transition Reasoning, and Sequential and Adaptive Reasoning. Guided by this taxonomy, we develop a systematic data pipeline and curate data mixtures spanning both pre-training and post-training.

To deliver strong physical-world understanding and interaction capabilities while supporting latency-sensitive deployment, we build our model on the Hy3-A3B language backbone and the Hy-ViT2 vision encoder. Its efficient Mixture-of-Experts architecture combines strong model capacity with high inference efficiency. We evaluate Hy-Embodied-VLM-1.0 on a comprehensive suite of 38 benchmarks covering embodied perception, physical-world understanding, and embodied reasoning. The model achieves the best performance among similarly sized models on 19 of the 38 benchmarks and substantially outperforms strong competitors, including Qwen3.6-A3B and Cosmos 3. Compared with the previous-generation Hy-Embodied-0.5 MoT-2B, Hy-Embodied-VLM-1.0 improves average performance by 8.4%. Despite activating only 3B parameters, it achieves performance close to that of the previous-generation model with 32B activated parameters. Beyond static benchmark evaluation, Hy-Embodied-VLM-1.0 also demonstrates strong performance on embodied agentic tasks requiring multi-turn interaction and long-horizon reasoning.

⭐️ Key Features

• 🧠 Efficient MoE, ~3B activated — Combines the Hy3-A3B language backbone with the Hy-ViT2 vision encoder in a Mixture-of-Experts architecture. Only ~3B parameters are activated per token — approximately one-tenth of the activated parameters of the previous-generation A32B system, while achieving nearly comparable overall performance.
• 🌏 Action-Centric Capability Taxonomy — We define three progressive levels of embodied intelligence: (i) Action-Relevant State Understanding for accurately understanding the states of the agent and its environment, (ii) Action–Transition Reasoning for understanding actions, planning them, and reasoning about their consequences, and (iii) Sequential and Adaptive Reasoning for long-horizon planning, reflection, repair, and recovery. Data and training are systematically designed around this taxonomy.
• 🔁 Self-Evolving Post-Training — Embodied agentic reasoning is cultivated through a self-evolving loop that couples reinforcement learning with rejection-sampling fine-tuning, seeded from a small curated set of high-quality thinking traces. A final reward-specialized stage trains continuous-reward and discrete-reward RL policies separately and fuses them, delivering sharp geometric precision alongside robust decision-making, planning, and reflection quality.
• 🏆 State-of-the-Art on Embodied Benchmarks — Ranks 1st on 19 of 38 benchmarks and 2nd on another 11, outperforming Qwen3.6-A3B (+4.4% avg), Cosmos 3-8B, and Embodied-R1.5-8B. State-of-the-art on R2R-CE vision-and-language navigation (RGB-only setting) and strong zero-shot performance on Matterport3D Object Goal Navigation.

🧱 Model Card

🛠️ Dependencies and Installation

Prerequisites

• 🖥️ Operating System: Linux (recommended)
• 🐍 Python: 3.10+
• ⚡ CUDA: 12.x (H100 / H20 / A100 tested)
• 🔥 PyTorch: 2.4+
• 🎮 GPU: NVIDIA GPU(s). The full BF16 model requires ~86 GB across GPUs; a single 8×80 GB node is sufficient.

Installation

We pin dependencies to the versions we validated end-to-end (vllm==0.14.1 + transformers==4.57.6 + torch==2.9.1). The cleanest way to install these — with the CUDA build matched to your driver — is uv:

# Install uv once (skip if you already have it)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Create a fresh venv (Python 3.10+)
uv venv --python 3.12
source .venv/bin/activate

# Clone the repo (provides the vLLM plugin and example scripts)
git clone https://github.com/Tencent-Hunyuan/HY-Embodied
cd HY-Embodied

🚀 Quick Start with vLLM (recommended)

vLLM is the recommended path for serving Hy-Embodied-VLM-1.0. Install vLLM together with matched torch/transformers wheels, then install this repo's plugin (registers the HYV3VL model and the reasoning / tool-call parsers):

# One-shot install: vllm + torch + torchvision + transformers at matching
# versions, with the CUDA build picked from your driver.
uv pip install vllm==0.14.1 --torch-backend auto

# Install this repo's vLLM plugin (registers HYV3VL model + parsers)
uv pip install -e Hy-Embodied-VLM-1.0/inference/vllm/

Start the server

serve.sh wraps vllm serve with the required flags (--reasoning-parser hunyuan_v3, --tool-call-parser hy_v3, --trust-remote-code, image mm-limit, chat template). It defaults to the Hub id tencent/Hy-Embodied-VLM-1.0, so no manual download is needed:

# TP=4 by default; override MODEL_PATH / TP / PORT via env vars.
bash Hy-Embodied-VLM-1.0/inference/vllm/serve.sh

Query the OpenAI-compatible endpoint

curl http://127.0.0.1:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hy_a3b",
    "messages": [{"role": "user", "content": "Describe how to grasp a cup."}],
    "max_tokens": 512,
    "chat_template_kwargs": {"enable_thinking": true}
  }'

Python (OpenAI SDK) — text, image, and streaming

import base64
from pathlib import Path
from openai import OpenAI  # pip install "openai>=1.30" pillow

client = OpenAI(base_url="http://127.0.0.1:8080/v1", api_key="EMPTY")

# --- text-only ---
resp = client.chat.completions.create(
    model="hy_a3b",
    messages=[{"role": "user", "content": "How do you open a fridge?"}],
    max_tokens=512,
    temperature=0.7,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)
msg = resp.choices[0].message
if getattr(msg, "reasoning_content", None):
    print("[thinking]", msg.reasoning_content)
print("[answer]  ", msg.content)

# --- image + text ---
def encode_image(path):
    mime = "image/jpeg" if path.lower().endswith((".jpg", ".jpeg")) else "image/png"
    return f"data:{mime};base64,{base64.b64encode(Path(path).read_bytes()).decode()}"

resp = client.chat.completions.create(
    model="hy_a3b",
    messages=[{
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": encode_image("example.jpg")}},
            {"type": "text", "text": "Describe the image in detail."},
        ],
    }],
    max_tokens=1024,
    temperature=0.7,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)
print(resp.choices[0].message.content)

See Hy-Embodied-VLM-1.0/inference/vllm/README.md for full serving options and example_client.py for the complete client (including streaming).

🤗 Alternative: HuggingFace transformers (single-instance)

For single-instance / offline inference without a server:

uv pip install torch==2.9.1 torchvision==0.24.1 --torch-backend auto
uv pip install transformers==4.57.6 accelerate pillow

# Run the demo (single image + a small batch; thinking and non-thinking modes)
cd Hy-Embodied-VLM-1.0/inference/transformers
python infer_hf.py

🧠 Reasoning-Mode Toggle

Hy-Embodied-VLM-1.0 is a hybrid reasoning model. Both modes are trained into the same weights; selection is per-request via a chat-template kwarg.

We deliberately use enable_thinking (Qwen3 convention) rather than reasoning_effort. vLLM prior to v0.22 has a top-level request.reasoning_effort field that silently clobbers chat_template_kwargs["reasoning_effort"] (fixed by vllm-project/vllm#43401); enable_thinking avoids the clobber and works across all vLLM versions.

🖥️ Hardware Requirements

• Full-precision inference: A single 8×80 GB GPU node (H100 / H20 / A100 80G). Model weights are BF16 (~86 GB); tensor-parallel size 4–8 recommended.
• Serving: 4 GPUs of 80 GB (tp=4) per replica is the recommended configuration for maximum throughput.
• Development / debugging: Any CUDA GPU. Smaller GPUs may require offloading or additional tensor parallelism.
• Disk: ~120 GB (including cache) for the model weights, auto-downloaded from the Hub on first run.

📊 Evaluation

Note: We evaluated Hy-Embodied-VLM-1.0 A3B across 38 embodied-relevant benchmarks against parameter-comparable state-of-the-art models. For detailed methodology, please refer to our technical report.

Action-Relevant State Understanding

Action–Transition Reasoning

Sequential and Adaptive Reasoning

Note: Hy-Embodied variants and Qwen3.6-A3B are evaluated in thinking mode; Embodied-R1.5-8B is only available in its Instruct configuration; Cosmos3-Nano-8B is reported in non-thinking mode (enabling thinking substantially degrades its performance).

📜 Older Versions

Prior releases of the Hy-Embodied family remain fully available:

📄 License

Released under Apache License 2.0. See LICENSE.

🏷️ Citation

If you find our work useful for your research and applications, please cite our tech reports using this BibTeX:

@article{tencent2026hyembodiedvlm10,
  title         = {Hy-Embodied-VLM-1.0: Efficient Physical-World Agents},
  author        = {Wang, Ziyi and Yu, Xumin and Rao, Yongming and Ling, Yonggen and Li, Yunheng and Wang, Oran and Gao, Mingqi and Zhou, Yuchen and Liang, Yves and Liu, Zuyan and others},
  year          = {2026},
  eprint        = {2607.12894},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2607.12894}
}

@article{tencent2026hyembodied05,
  title         = {HY-Embodied-0.5: Embodied Foundation Models for Real-World Agents},
  author        = {Team, HY and Yu, Xumin and Liu, Zuyan and Wang, Ziyi and Zhang, He and Rao, Yongming and Liu, Fangfu and Zhang, Yani and Zhao, Ruowen and Wang, Oran and others},
  year          = {2026},
  eprint        = {2604.07430},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2604.07430}
}

🙏 Acknowledgements

Built on the Hy3 MoE LLM backbone and the Hy-ViT2 vision encoder. We thank the broader Tencent Hunyuan and Robotics X communities for infrastructure, evaluation resources, and design feedback.