Open source robot brain: 22 benchmark 16 first, against Gemini
🤖 Intelling open-source robot brain: new smart player
Zeng Tse Yuan was released yesterday. HY-Embodied-0.5It's not an ordinary big language model -- it's for robots.BrainI don't know. And...Full Open SourceI don't know.
This time, it's really serious.
Two versions, two strokes.
HY-Embodied-0.5 provides two versions:
| Version | Parameters | Positioning | Performance |
|---|---|---|---|
| MoT-2B | 2B Activate / 4B Total Parameters | Marginal deployment (robots) | 22 Baseline test, 16 crushing peer models |
| 32B | 32B Parameters | Complex reasoning (cloud/server) | Contrast Gemini 3.0 Pro |
MoT-2B highlights: Although the overall parameter is 4B, only 2.2B parameters are activated in reasoning. Which means...The speed of a 2B model with the ability to exceed 2BI don't know.
Four core technologies
1. MoT (Mixture-of-Transformers) architecture
Traditional VLM (visual language model) combines visual and language. HY-Embodied usedMoT Architecture, and achieves the specific calculation of the modulus by latent tokens:
- Visual PathSpecialized visual treatment module
- Language PathSpecialized text processing module
- Synergical reasoning: Joint reasoning across the paradigm
Effects of this:The visual sensory ability has improved dramatically., while maintaining an efficient pace of reasoning.
2. Chain reasoning + evolutive distillation
How can a small model have a large model of reasoning? The answer is...Distillation:
- 32B Models learn complex step-by-step reasoning, planning, thinking
- Through on-policy distillation,Move to 2B Model
- Small models can be made.Chain reasoningOne step at a time.
3. Large-scale pre-training
Data size: Over 100 million Physical and spatial data points, over 200 billion Tokens.
That's what makes models learn:
- 3D Space Deep Understanding
- Interactive approach of physical objects
- Agent's Dynamics
It's not just text data., the data specifically for robotic scenarios.
4. VLA (Vision-Language-Action) application
HY-Embodied is not just a model for running benchmark tests, it's designed to beThe core cognitive engine of a physical robot.:
- A seamless integration into a VLA framework
- As robot brain drives complex tasks.
- The high success rate of real world robots.
🎯 Baseline test: crushing peer
In 22 smart-relevant benchmark tests, MoT-2B is16 crushed peer models:
| Benchmark | HY-Embodied | Qwen3-VL 2B | RoboBrain 4B |
|---|---|---|---|
| CV-Bench | 89.2 | 80.0 | 86.9 |
| DA-2K | 92.3 | 69.5 | 79.4 |
| EmbSpatial-Bench | 82.8 | 75.9 | 73.8 |
| MindCube | 66.3 | 28.4 | 26.9 |
The most obvious difference is MindCube.:HY-Embodied 66.3 vs Qwen3-VL 28.4,2.3 Multiple advantagesI don't know.
Why is this important?
Smart.AI. Key tracks for the next decade.:
- Language Model: solved the problem of “understanding” and “generation”
- Visual Model: solved the "see" problem
- Smart.: Resolve "understanding + see + action" full closed loop
Robots need:
- Perception3D Space, object position, environmental change
- Arguments: plan the course of action, predict the consequences
- Implementation: Control of electrical power, physical operation
HY-Embodied solvedFirst two questionsAnd it gives the robot a real brain.
How does it work?
Open source address:
- GitHub:https://github.com/Tencent-Hunyuan/HY-Embodied
- Hugging Face:https://huggingface.co/tencent/HY-Embodied-0.5
# 安装 pip install git+https://github.com/huggingface/transformers@... # 克隆仓库 git clone https://github.com/Tencent-Hunyuan/HY-Embodied cd HY-Embodied/ pip install -r requirements.txt # 运行推理 python inference.py
Hardware requirements:
- GPU: at least 16GB VRAM (recommended NVIDIA GPU)
- CPU: Supported but slow
- Storage: 20GB+ (model about 8GB)
Summary
Open source of HY-Embodied-0.5, markingSmart enough to enter the Open Age:
- 2B version crushing the same level, suitable for deployment at the edge of a robot.
- 32B version pairing Gemini 3.0Pro, suitable for cloud reasoning
- MoT architecture + chain reasoning + 100M+ data training
- VLA applications give robots a real brain.
This is...A major breakthrough for Chinese firms in the area of intelligence.I don't know.
The robot has a brain, and it's going to work.