Skip to content
embodied.guide

BeingBeyondChina2026

Being-H0.5

Runs onAdam-Uby PNDbotics AG1by Unitree Robotics A

Key figures as stated by the developer

7 of 7 key figures stated
  • Model type ⁠B vision-language-action
  • Controls ⁠A arms and hands
  • Inputs ⁠A camera language robot state
  • Access ⁠C Open weights (download)
  • Parameters ⁠B 2B
  • Base model ⁠A InternVL-3.5
  • Licence ⁠A Free, commercial use

Full specifications

13 of 14 fields stated
The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field Value Source and caveats
Release
Released19 January 2026

[A] "Submitted on 19 Jan 2026" ⁠D
Input and output
Model typevision-language-action

BeingBeyond's own term: "Base vision-language-action model (preview)". ⁠B

Inputscamera language robot state

Technical report, LIBERO setup: "The policy takes multi-view RGB observations (wrist-mounted and third-person cameras)". Robot state: "concatenating proprioceptive state embeddings and current action-token embeddings". ⁠A

Controlsarms and hands

Technical report: "We evaluate on five robotic platforms spanning upper-body humanoids, single-arm dexterous manipulation, and lightweight parallel grippers". ⁠A

Degrees of freedom controllednot stated
Size and speed
Parameters2B

Repository README, checkpoint table: "| 2B | Base vision-language-action model (preview)". ⁠B

Control rate50Hz

Technical report, platform control rates the real-time stack serves: "10 Hz tabletop manipulators to 50 Hz humanoids". ⁠A
Where it runs and what it is built on
On robot or cloudboth

Technical report: one checkpoint reaches real-time performance "under both cloud computing (high network latency) and edge deployment". ⁠A

Base modelInternVL-3.5

Technical report: "The backbone of Being-H0.5 is initialized from InternVL-3.5". ⁠A
Openness
AccessOpen weights (download)

Model card: "Download models from Hugging Face". ⁠C

Licence termsApache-2.0

Model card header: "license: apache-2.0". ⁠C

LicenceFree, commercial use

Technical report: "we provide a comprehensive open-source release encompassing our model weights, training pipeline, and simulation scripts". Weights licence: "license: apache-2.0". ⁠A
The developer’s own claims
Learns new tasksneeds retraining

Developer's own checkpoints for benchmarks are fine-tuned versions: "Post-trained on LIBERO benchmark". ⁠B

Works across robotsyes

Developer claim: "We empirically deploy a single checkpoint of Being-H0.5 to control PND Adam-U, Franka+Inspire, Unitree G1, BeingBeyond D1, and LeRobot SO-101". ⁠A

Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.
How to read the figures
50HzA
The letter in the grey square after a figure points to the source the figure came from. Pages with several sources carry several letters.

What the data labels mean

Notes

  • BeingBeyond publishes a 2B-parameter pretrained checkpoint and task-specific checkpoints on Hugging Face; the table marks the base one as a preview.
  • BeingBeyond lists its Beijing address on its own website.
Versions and notes in the key figures (7)

Model type

BeingBeyond's own term: "Base vision-language-action model (preview)".

Controls

Technical report: "We evaluate on five robotic platforms spanning upper-body humanoids, single-arm dexterous manipulation, and lightweight parallel grippers".

Inputs

Technical report, LIBERO setup: "The policy takes multi-view RGB observations (wrist-mounted and third-person cameras)". Robot state: "concatenating proprioceptive state embeddings and current action-token embeddings".

Access

Model card: "Download models from Hugging Face".

Parameters

Repository README, checkpoint table: "| 2B | Base vision-language-action model (preview)".

Base model

Technical report: "The backbone of Being-H0.5 is initialized from InternVL-3.5".

Licence

Technical report: "we provide a comprehensive open-source release encompassing our model weights, training pipeline, and simulation scripts". Weights licence: "license: apache-2.0".

Sources

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.