Hy-Embodied-VLM-1.0

An efficient vision-language model for embodied agents working in the physical world, activating about 3B of its roughly 30B parameters per token.

Made by Tencent Hunyuan.

Licenses

Code: Apache-2.0 (opens in a new tab)
Read by Robotheca from LICENSE (opens in a new tab) (Apache-2.0) at commit 16289e9 (opens in a new tab), on 16 Oct 2026.
Weights: Apache-2.0 (opens in a new tab)
Read on huggingface.co/tencent/Hy-Embodied-VLM-1.0/blob/main/LICENSE (opens in a new tab).
Commercial use: Yes

Mixture-of-Experts model combining a Hy3-A3B language backbone with the Hy-ViT2 vision encoder.
Built around an action-centric taxonomy: state understanding, action-transition reasoning, and sequential adaptive reasoning.
Supports a hybrid reasoning mode selectable per request, with chain-of-thought or direct answers.
Post-training uses a self-evolving loop combining reinforcement learning and rejection-sampling fine-tuning.
Includes a navigation agent demo for instruction-following and object-goal navigation.
Inference is available through HuggingFace transformers or an OpenAI-compatible vLLM server; fine-tuning code is not yet released.

Written by Robotheca from the project's README (opens in a new tab) at commit 16289e9.

Difficulty: Advanced
Hardware it needs, as the project states it: Linux (recommended) with Python 3.10+ and CUDA 12.x (H100, H20 and A100 tested); the full BF16 model needs about 86 GB across GPUs, and a single 8×80 GB node is sufficient; 4 GPUs of 80 GB (tp=4) per replica is recommended for serving; any CUDA GPU for development, with smaller GPUs possibly needing
How to start: The project's README (opens in a new tab)

Checked on GitHub on 16 Oct 2026.

No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.

Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.