tencent / HY-Embodied-0.5

🔥 Updates

• [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

We introduce HY-Embodied-0.5 , a suite of foundation models tailored specifically for real-world embodied intelligence. To bridge the gap between general Vision-Language Models (VLMs) and the strict demands of physical agents, our models are engineered to excel in spatial-temporal visual perception and complex embodied reasoning (prediction, interaction, and planning).

The suite features an innovative Mixture-of-Transformers (MoT) architecture utilizing latent tokens for modality-specific computing, significantly enhancing fine-grained perception. It includes two primary variants: a highly efficient 2B model for edge deployment and a powerful 32B model for complex reasoning. Through a self-evolving post-training paradigm and large-to-small on-policy distillation, our compact MoT-2B outperforms state-of-the-art models of similar size across 16 benchmarks, while the 32B variant achieves frontier-level performance comparable to Gemini 3.0 Pro. Ultimately, HY-Embodied serves as a robust "brain" for Vision-Language-Action (VLA) pipelines, delivering compelling results in real-world physical robot control.

⭐️ Key Features

• 🧠 Evolved MoT Architecture: Designed for maximum efficiency without sacrificing visual acuity. The MoT-2B variant contains 4B total parameters but requires only 2.2B activated parameters during inference. By emphasizing modality-specific computing in the vision pathway, it achieves the high inference speed of a dense 2B model while delivering superior, fine-grained perceptual representations.
• 🔗 High-Quality Mixed Chain Reasoning: We introduce an advanced iterative, self-evolving post-training pipeline. By employing on-policy distillation, we successfully transfer the sophisticated step-by-step reasoning, planning, and high-quality "thinking" capabilities from our powerful 32B model directly to the compact 2B variant.
• 🌍 Large-Scale Embodied Pre-training: Grounded in a massive, specially curated dataset comprising >100 million embodied and spatial-specific data points. Trained on a corpus exceeding 200 billion tokens , the model develops a deep, native understanding of 3D spaces, physical object interactions, and agent dynamics.
• 🦾 Stronger VLA Application: Beyond standard academic benchmarks, HY-Embodied is engineered to be the core cognitive engine for physical robots. It seamlessly integrates into Vision-Language-Action (VLA) frameworks, acting as a highly robust and capable brain to drive high success rates in complex, real-world robotic control tasks.

📅 Plannings

• Transformers Inference
• vLLM Inference
• Online Gradio Demo

🛠️ Dependencies and Installation

Prerequisites

• 🖥️ Operating System : Linux (recommended)
• 🐍 Python : 3.12+ (recommended and tested)
• ⚡ CUDA : 12.6
• 🔥 PyTorch : 2.8.0
• 🎮 GPU : NVIDIA GPU with CUDA support

Installation

1. Install the specific Transformers version required for this model:

pip install git+https://github.com/huggingface/transformers@9293856c419762ebf98fbe2bd9440f9ce7069f1a

Note : We will merge the improvements into the Transformers main branch later.

1. Install other dependencies:

pip install -r requirements.txt

Quick Start

1. Clone the repository:

git clone https://github.com/Tencent-Hunyuan/HY-Embodied
cd HY-Embodied/

1. Install dependencies:

pip install -r requirements.txt

1. Run inference:

python inference.py

The example script demonstrates both single generation and batch generation capabilities.

Model Download

The code automatically downloads the model tencent/HY-Embodied-0.5 from Hugging Face Hub. Ensure you have sufficient disk space (8 GB) for the model weights.

Hardware Requirements

• GPU : Recommended for optimal performance (NVIDIA GPU with at least 16GB VRAM)
• CPU : Supported but slower
• Memory : At least 16GB RAM recommended
• Storage : 20GB+ free space for model and dependencies

🚀 Quick Start with Transformers

Basic Inference Example

import os
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

# Load model & processor
MODEL_PATH = "tencent/HY-Embodied-0.5"
DEVICE = "cuda"
THINKING_MODE = False
TEMPERATURE = 0

processor = AutoProcessor.from_pretrained(MODEL_PATH)

# Load chat template if available
chat_template_path = os.path.join(MODEL_PATH, "chat_template.jinja")
if os.path.exists(chat_template_path):
    processor.chat_template = open(chat_template_path).read()

model = AutoModelForImageTextToText.from_pretrained(MODEL_PATH, torch_dtype=torch.bfloat16)
model.to(DEVICE).eval()

# Prepare input messages
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "./figures/example.jpg"},
            {"type": "text", "text": "Describe the image in detail."},
        ],
    }
]

# Process and generate
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
    enable_thinking=THINKING_MODE,
).to(model.device)

with torch.no_grad():
    generated_ids = model.generate(
        **inputs,
        max_new_tokens=32768,
        use_cache=True,
        temperature=TEMPERATURE,
        do_sample=TEMPERATURE > 0,
    )

output_ids = [out[len(inp):] for inp, out in zip(inputs.input_ids, generated_ids)]
print(processor.batch_decode(output_ids, skip_special_tokens=True)[0])

Batch Inference

import os
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

# Load model & processor
MODEL_PATH = "tencent/HY-Embodied-0.5"
DEVICE = "cuda"
THINKING_MODE = False
TEMPERATURE = 0

processor = AutoProcessor.from_pretrained(MODEL_PATH)

# Load chat template if available
chat_template_path = os.path.join(MODEL_PATH, "chat_template.jinja")
if os.path.exists(chat_template_path):
    processor.chat_template = open(chat_template_path).read()

model = AutoModelForImageTextToText.from_pretrained(MODEL_PATH, torch_dtype=torch.bfloat16)
model.to(DEVICE).eval()

# Batch Inference (multiple prompts at once)
messages_batch = [
    # Sample A: image + text
    [
        {
            "role": "user",
            "content": [
                {"type": "image", "image": "./figures/example.jpg"},
                {"type": "text", "text": "Describe the image in detail."},
            ],
        }
    ],
    # Sample B: text only
    [
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "How to open a fridge?"},
            ],
        }
    ],
]

# Process each message independently
all_inputs = []
for msgs in messages_batch:
    inp = processor.apply_chat_template(
        msgs,
        tokenize=True,
        add_generation_prompt=True,
        return_dict=True,
        return_tensors="pt",
        enable_thinking=THINKING_MODE,
    )
    all_inputs.append(inp)

# Left-pad and batch
batch = processor.pad(all_inputs, padding=True, padding_side="left").to(model.device)

with torch.no_grad():
    batch_generated_ids = model.generate(
        **batch,
        max_new_tokens=32768,
        use_cache=True,
        temperature=TEMPERATURE,
        do_sample=TEMPERATURE > 0,
    )

# Decode: strip the padded input portion
padded_input_len = batch["input_ids"].shape[1]
for i, msgs in enumerate(messages_batch):
    out_ids = batch_generated_ids[i][padded_input_len:]
    print(f"\n--- Sample {i} ---")
    print(processor.decode(out_ids, skip_special_tokens=True))

📊 Evaluation

Visual Perception

Note : We evaluated HY-Embodied-0.5 MoT-2B across 22 embodied-relevant benchmarks against models of similar size. For detailed performance metrics and methodology, please refer to our technical report.

Note : We observed that small models from the Qwen3.5 series produce repetitive thinking patterns in some benchmarks, which leads to lower overall results. Therefore, we compare against Qwen3-VL models in our evaluations.

Embodied Understanding

Spatial Understanding

Note: Results for HY-Embodied-0.5 MoT-2B are reported in thinking mode, while for all other models, we report the better performance between non-thinking and thinking modes.

📚 Citation

Citation information will be available upon publication. Please check back for updates.

🙏 Acknowledgements

We thank the Hugging Face community for their support and the open-source contributions that made this implementation possible.