tencent/HY-Embodied-0.5 · Hugging Face

来源：https://huggingface.co/tencent/HY-Embodied-0.5 面向真实世界智能体的具身基础模型系列

腾讯Robotics X × HY视觉团队

技术报告 (https://github.com/Tencent-Hunyuan/HY-Embodied/blob/master/hy_embodied_tech_report.pdf)论文 (https://arxiv.org/abs/2604.07430)模型 (https://huggingface.co/tencent/HY-Embodied-0.5/tree/main)GitHub (https://github.com/Tencent-Hunyuan/HY-Embodied)X (https://x.com/TencentHunyuan/status/2042503238877135336?s=20)

https://huggingface.co/tencent/HY-Embodied-0.5#%F0%9F%94%A5-updates🔥 更新

• [2026\-04\-09]🚀 我们已发布HY-Embodied-0.5，开源了HY\-Embodied\-0\.5 MoT\-2B的权重，可在Hugging Face (https://huggingface.co/tencent/HY-Embodied-0.5/tree/main)获取，并附带官方推理代码！

https://huggingface.co/tencent/HY-Embodied-0.5#%F0%9F%93%96-abstract📖 摘要

我们推出HY-Embodied-0.5，一套专为真实世界具身智能量身定制的基础模型系列。为弥合通用视觉语言模型（VLM）与物理智能体严格需求之间的差距，我们的模型在时空视觉感知和复杂具身推理（预测、交互和规划）方面表现出色。

该系列采用创新的Mixture-of-Transformers（MoT）架构，利用潜在令牌实现模态特定计算，显著增强细粒度感知能力。它包含两个主要变体：一个用于边缘部署的高效2B模型和一个用于复杂推理的强大32B模型。通过自我进化的后训练范式和大到小的在策略蒸馏，我们紧凑的MoT-2B在16项基准测试中超越了同类尺寸的最先进模型，而32B变体则达到了与Gemini 3.0 Pro相当的前沿水平。最终，HY-Embodied作为视觉-语言-动作（VLA）管线的强大“大脑”，在真实世界机器人控制中展现出令人瞩目的效果。

HY-Embodied 预告

https://huggingface.co/tencent/HY-Embodied-0.5#%E2%AD%90%EF%B8%8F-key-features⭐️ 关键特性

• 🧠进化版MoT架构：在保持视觉敏锐度的同时追求极致效率。MoT-2B变体总参数量为4B，但在推理时仅需2.2B的激活参数。通过在视觉通路中强调模态特定计算，它实现了与密集2B模型相当的高推理速度，同时提供了更优越的细粒度感知表示。
• 🔗**高质量混合链式推理：**我们引入了一种先进的迭代、自我进化后训练管线。通过在策略蒸馏，我们将强大32B模型的复杂逐步推理、规划和高质量“思考”能力成功迁移到紧凑的2B变体上。
• 🌍大规模具身预训练：基于一个包含超过1亿具身和空间特定数据点的大规模精心策划的数据集。在超过2000亿令牌的语料上进行训练，模型对3D空间、物理对象交互和智能体动力学形成了深入的原生理解。
• 🦾**更强的VLA应用：**除了标准的学术基准，HY-Embodied被设计为物理机器人的核心认知引擎。它无缝集成到视觉-语言-动作（VLA）框架中，作为高度鲁棒且能力强大的大脑，在复杂的真实世界机器人控制任务中实现高成功率。

HY-Embodied 架构

https://huggingface.co/tencent/HY-Embodied-0.5#%F0%9F%93%85-plannings📅 计划

• Transformers 推理
• vLLM 推理
• 微调代码
• 在线 Gradio 演示

https://huggingface.co/tencent/HY-Embodied-0.5#%F0%9F%9B%A0%EF%B8%8F-dependencies-and-installation🛠️ 依赖与安装

https://huggingface.co/tencent/HY-Embodied-0.5#prerequisites先决条件

• 🖥️操作系统：Linux（推荐）
• 🐍Python：3.12+（推荐并已测试）
• ⚡CUDA：12.6
• 🔥PyTorch：2.8.0
• 🎮GPU：支持CUDA的NVIDIA GPU

https://huggingface.co/tencent/HY-Embodied-0.5#installation安装

1. 安装此模型所需的特定Transformers版本：

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

注意：后续我们会将改进合并到Transformers主分支中。

2. 安装其他依赖：

pip install -r requirements.txt

https://huggingface.co/tencent/HY-Embodied-0.5#quick-start快速开始

1. 克隆仓库：

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

2. 安装依赖：

pip install -r requirements.txt

3. 运行推理：

python inference.py

示例脚本演示了单次生成和批量生成功能。

https://huggingface.co/tencent/HY-Embodied-0.5#model-download模型下载

代码会自动从Hugging Face Hub下载模型tencent/HY\-Embodied\-0\.5。请确保有足够的磁盘空间（8 GB）存放模型权重。

https://huggingface.co/tencent/HY-Embodied-0.5#hardware-requirements硬件要求

• GPU：推荐用于最佳性能（至少16GB显存的NVIDIA GPU）
• CPU：支持但速度较慢
• 内存：建议至少16GB RAM
• 存储：模型和依赖需要20GB+空闲空间

https://huggingface.co/tencent/HY-Embodied-0.5#%F0%9F%9A%80-quick-start-with-transformers🚀 使用 Transformers 快速开始

https://huggingface.co/tencent/HY-Embodied-0.5#basic-inference-example基础推理示例

`` import os import torch from transformers import AutoModelForImageTextToText, AutoProcessor

加载模型和处理器

MODEL_PATH = “tencent/HY-Embodied-0.5” DEVICE = “cuda” THINKING_MODE = False TEMPERATURE = 0.8

processor = AutoProcessor.from_pretrained(MODEL_PATH)

如果存在聊天模板则加载

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()

准备输入消息

messages = [ { “role”: “user”, “content”: [ {“type”: “image”, “image”: “./figures/example.jpg”}, {“type”: “text”, “text”: “Describe the image in detail.”}, ], } ]

处理并生成

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]) ``

https://huggingface.co/tencent/HY-Embodied-0.5#batch-inference批量推理

`` import os import torch from transformers import AutoModelForImageTextToText, AutoProcessor

加载模型和处理器

MODEL_PATH = “tencent/HY-Embodied-0.5” DEVICE = “cuda” THINKING_MODE = False TEMPERATURE = 0.8

processor = AutoProcessor.from_pretrained(MODEL_PATH)

如果存在聊天模板则加载

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()

批量推理（同时处理多个提示）

messages_batch = [ # 样本 A: 图像 + 文本 [ { “role”: “user”, “content”: [ {“type”: “image”, “image”: “./figures/example.jpg”}, {“type”: “text”, “text”: “Describe the image in detail.”}, ], } ], # 样本 B: 仅文本 [ { “role”: “user”, “content”: [ {“type”: “text”, “text”: “How to open a fridge?”}, ], } ], ]

独立处理每条消息

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)

左侧填充并批量处理

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, )

解码：去除填充的输入部分

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)) ``

https://huggingface.co/tencent/HY-Embodied-0.5#%F0%9F%93%8A-evaluation📊 评估

https://huggingface.co/tencent/HY-Embodied-0.5#visual-perception视觉感知

注意：我们在22项具身相关基准上，将HY-Embodied-0.5 MoT-2B与相似尺寸的模型进行了比较。详细的性能指标和方法论请参考我们的技术报告。

注意：我们观察到Qwen3.5系列的小模型在某些基准上产生重复思考模式，导致整体结果较低。因此在评估中，我们与Qwen3-VL模型进行对比。

https://huggingface.co/tencent/HY-Embodied-0.5#embodied-understanding具身理解

https://huggingface.co/tencent/HY-Embodied-0.5#spatial-understanding空间理解

注意：HY-Embodied-0.5 MoT-2B的结果为思考模式下的表现，而其他所有模型，我们报告的是非思考模式与思考模式中更好的表现。

https://huggingface.co/tencent/HY-Embodied-0.5#%F0%9F%93%9A-citation📚 引用

如果您觉得本研究对您的研究和应用有帮助，请使用以下BibTeX引用我们的论文：

@article{tencent2026hyembodied05, title={HY-Embodied-0.5: Embodied Foundation Models for Real-World Agents}, author={Tencent Robotics X and HY Vision Team}, journal={arXiv preprint arXiv:2604.07430}, year={2026} }

https://huggingface.co/tencent/HY-Embodied-0.5#%F0%9F%99%8F-acknowledgements🙏 致谢

我们感谢Hugging Face社区的支持以及开源贡献，使这一实现成为可能。