Tencent has released Hy-Embodied-VLM-1.0, an efficient Mixture-of-Experts vision–language foundation model designed for embodied agents operating in the physical world. The model activates only approximately 3 billion parameters per token out of a total of 30 billion, aiming to balance high inference efficiency with strong physical-world understanding.
- Built on the Hy3-A3B language backbone and Hy-ViT2 vision encoder within an efficient MoE architecture.
- Achieves state-of-the-art performance on 19 of 38 benchmarks covering embodied perception and reasoning, outperforming Qwen3.6-A3B and Cosmos 3.
- Improves average performance by 8.4% compared to the previous-generation Hy-Embodied-0.5 MoT-2B model.
- Utilizes a self-evolving post-training loop coupling reinforcement learning with rejection-sampling fine-tuning.
- Weights and inference code for Hugging Face transformers and vLLM are available under the Apache 2.0 license.
The release provides open-source weights and tools to support latency-sensitive deployment of embodied agents requiring multi-turn interaction and long-horizon reasoning.