<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> HY-Embodied-VLM-1.0 is an efficient vision-language foundation model designed for embodied agents operating in the physical world. It addresses multimodal perception, action reasoning, and adaptation needs for robotic agents, offering high inference efficiency via a Mixture-of-Experts (MoE) architecture.

How It Works

This model utilizes an efficient MoE architecture, integrating the Hy3-A3B language backbone with the Hy-ViT2 vision encoder. It activates only ~3 billion parameters per token out of ~30 billion total, enhancing inference speed. Development is guided by an action-centric capability taxonomy (state understanding, action-transition reasoning, sequential/adaptive reasoning), supported by a systematic data pipeline and training strategy. A self-evolving post-training loop, combining RL and rejection-sampling fine-tuning, cultivates embodied agentic reasoning.

Quick Start & Requirements

• Installation: Primarily via uv for dependency management, requiring Python 3.10+. Options include vLLM (serving) or HuggingFace transformers (single-instance).
• Prerequisites: Linux (recommended), Python 3.10+, CUDA 12.x, PyTorch 2.4+, NVIDIA GPU(s).
• Resource Footprint: Full BF16 model needs ~86 GB across GPUs (8x80 GB node sufficient for training/inference). Serving recommends 4x 80 GB GPUs/replica. ~120 GB disk space for weights.
• Links: Inference code for HuggingFace transformers and vLLM available. Model weights on Hugging Face.

Highlighted Details

• Achieves state-of-the-art on 19/38 embodied benchmarks, outperforming Qwen3.6-A3B and Cosmos 3.
• Efficient MoE activates ~3B parameters/token, matching previous 32B models' performance.
• Powers closed-loop navigation agents with strong zero-shot performance on R2R-CE and Matterport3D.
• Hybrid reasoning model supports complex planning (enable_thinking=True) or direct answers (enable_thinking=False).

Maintenance & Community

Recent updates include Hy-Embodied-VLM-1.0 (July 2026) and HY-VLA-0.5 (June 2026). Developed by Tencent Robotics X, Hy Vision Team, and Futian Laboratory. No specific community channels listed.

Licensing & Compatibility

Released under the Apache License 2.0, which is permissive for commercial use and closed-source linking.

Limitations & Caveats

Fine-tuning code is not yet released. High hardware requirements (~86 GB BF16 model) limit deployment on consumer hardware. Cosmos3-Nano-8B performance degrades significantly when thinking is enabled.