Skip to content
embodied.guide

Alibaba QwenChina2026

Qwen-VLA

Key figures as stated by the developer

4 of 7 key figures stated
  • Model type ⁠B vision-language-action
  • Controls ⁠B arms and hands
  • Inputs ⁠A camera language
  • Base model ⁠B Qwen3.5-4B (vision-language backbone)

Not stated: Access · Parameters · Licence

Full specifications

7 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
Released28 May 2026

[A] "Submitted on 28 May 2026" ⁠C
Input and output
Model typevision-language-action

Developer wording, GitHub README: "a unified vision-language-action generalist model". ⁠B

Inputscamera language

Technical report: camera views are labelled per source, "a head-mounted ego-centric camera alongside one or two wrist-mounted cameras"; the model generalises "to generalise to a broader range of natural-language instructions at inference time". Robot state was tested as an extra input only in an ablation. ⁠A

Controlsarms and hands

README: "Manipulation, navigation, egocentric action modeling, and trajectory prediction share one action-and-trajectory prediction space." The model also predicts navigation waypoints. ⁠B

Degrees of freedom controllednot stated

Note The count differs per robot: "Different robot platforms have different action dimensionalities and semantics". ⁠A
Size and speed
Parametersnot stated

Note The developer gives parts, not a total: a Qwen3.5-4B backbone and a "1.15B DiT flow-matching action decoder". ⁠B

Control ratenot stated

Note The rate is set per robot in the prompt: the prompt values "reflect the dataset’s original control frequency". ⁠A
Where it runs and what it is built on
On robot or cloudnot stated
Base modelQwen3.5-4B (vision-language backbone)

README: Qwen3.5-4B backbone with a "1.15B DiT flow-matching action decoder". ⁠B
Openness
Accessnot stated
Licence termsnot stated
Licencenot stated
The developer’s own claims
Learns new tasksneeds retraining

Developer claim: the real-robot results come "by fine-tuning on demonstrations collected through ALOHA across multiple tasks". ⁠A

Works across robotsyes

Developer claim, README: "One set of weights serves multiple platforms; switching embodiments requires only changing a text prompt." ⁠B

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

  • The Qwen team describes Qwen-VLA as a unified vision-language-action generalist model.
  • The Qwen-VLA technical report was submitted to arXiv on 28 May 2026.
Versions and notes in the key figures (5)

Model type

Developer wording, GitHub README: "a unified vision-language-action generalist model".

Controls

README: "Manipulation, navigation, egocentric action modeling, and trajectory prediction share one action-and-trajectory prediction space." The model also predicts navigation waypoints.

Inputs

Technical report: camera views are labelled per source, "a head-mounted ego-centric camera alongside one or two wrist-mounted cameras"; the model generalises "to generalise to a broader range of natural-language instructions at inference time". Robot state was tested as an extra input only in an ablation.

Parameters

Note The developer gives parts, not a total: a Qwen3.5-4B backbone and a "1.15B DiT flow-matching action decoder".

Base model

README: Qwen3.5-4B backbone with a "1.15B DiT flow-matching action decoder".

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.