Overview
Hy-Embodied-0.5-VLA-UMI is Tencent’s pre-trained vision-language-action (VLA) checkpoint for robotic manipulation. It is a general-purpose starting point for fine-tuning on target robots, rather than the separately released RoboTwin post-trained checkpoint. The official release was announced on June 15, 2026. (github.com)
Capabilities and uses
- Combines visual observations and language instructions to generate continuous robot action chunks for manipulation tasks.
- Uses the Hy-Embodied-0.5 MoT vision-language backbone, a 370-million-parameter flow-matching action expert and a compact memory encoder. The UMI checkpoint was pre-trained in single-frame mode; the memory encoder is activated during supervised fine-tuning.
- Pre-trained on more than 10,000 hours of UMI demonstrations, covering approximately one million episodes and over 70 tasks. Its action representation uses relative end-effector motion, with a 50-step horizon at 10 Hz.
- Intended for research and adaptation to downstream robot embodiments and tasks. (huggingface.co)
Limitations and availability
- The UMI checkpoint is designed for fine-tuning; successful deployment requires adaptation and validation for the target robot and environment. The reported RoboTwin 2.0 scores concern the evaluated Hy-VLA system, not a guarantee of performance from the untouched UMI checkpoint.
- The official repository provides code and model weights under Apache 2.0. The published inference example uses CUDA and bfloat16; local deployment has associated hardware and operating costs. (github.com)
Pricing
As of November 16, 2026, model-specific hosted API input and output prices could not be verified in either Tencent’s mainland China CNY pricing documentation or its international USD pricing documentation. No billing unit or amount is confirmed for this checkpoint; prices for other models should not be applied to it. Consult the official mainland pricing page and international pricing page for updates. (cloud.tencent.com)