Overview
Released on July 15, 2026, Hy-Embodied-VLM-1.0 is an open-weight vision-language foundation model for embodied agents operating in the physical world. It was developed by Tencent Robotics X, the Hy Vision Team, and Futian Laboratory. (github.com)
Capabilities and uses
- Understands action-relevant physical and spatial states, reasons about actions and their consequences, and supports sequential planning, reflection, and recovery.
- Intended for embodied-agent research and applications such as scene understanding, action planning, and vision-language navigation.
- Uses a mixture-of-experts architecture with approximately 30 billion total parameters and 3 billion activated per token. It accepts text and images, with up to 128 images per prompt and a 32,768-token context window. Thinking and non-thinking modes are supported. (huggingface.co)
- In the developers' evaluation, it ranked first among comparable models on 19 of 38 embodied benchmarks. These results are benchmark-specific, not a guarantee of real-world performance. (arxiv.org)
Availability and limitations
Weights and inference code are available for self-hosting under Apache License 2.0. The official documentation recommends vLLM for serving and also provides a Transformers inference path. Full BF16 weights require approximately 86 GB across GPUs; the documented serving recommendation is four 80 GB GPUs per replica. Smaller GPUs may require offloading or additional tensor parallelism. Outputs and action plans should be validated before use with physical systems; the VLM is not by itself a complete robot-control system. (huggingface.co)
Pricing
As of November 16, 2026, model-specific official hosted API input and output prices could not be verified in either CNY on Tencent Cloud's domestic pricing page or USD on its international pricing page. Self-hosting may incur infrastructure costs; no per-token API price is inferred from the open-weight release. (cloud.tencent.com)
Official sources
- Announcement and code: https://github.com/Tencent-Hunyuan/HY-Embodied
- Model card and weights: https://huggingface.co/tencent/Hy-Embodied-VLM-1.0
- Domestic pricing page, checked November 16, 2026: https://cloud.tencent.com/document/product/1823/130055/
- International pricing page, checked November 16, 2026: https://intl.cloud.tencent.com/ind/document/product/1300/78937