Alibaba DAMO Academy · robot AI model
RynnVLA-Latent
China · 2026
Runs on S1 by AstribotSource and note: S1B arxiv.org/html/2610.34356 · Retrieved 30 October 2026
Sources retrieved: 30 October 2026 (A, B, C, D)
- Model type
- Vision-language-action (VLA)Source and note: Model typeA huggingface.co/Alibaba-DAMO-Academy/RynnVLA-Latent-4B · Retrieved 30 October 2026Note. Alibaba's own term: “a Mixture-of-Transformers vision-language-action model”.
- What it controls
- Arms and handsSource and note: What it controlsB arxiv.org/html/2610.34356 · Retrieved 30 October 2026Note. Technical report, after robot post-training: “The policy predicts native action chunks, which are passed to the robot control interface for execution.” The real-robot platforms are dual-arm: “Two dual-arm embodiments used for real-world evaluation”. The report does not say whether the mobile base is driven by the policy.
- Inputs
- Camera Language Robot stateSource and note: InputsA huggingface.co/Alibaba-DAMO-Academy/RynnVLA-Latent-4B · Retrieved 30 October 2026Note. Model card: “Given a language instruction and up to six camera views”. Robot state is added in robot post-training: “Proprioceptive state is introduced only at this stage.”
- Access
- Open weights (download)Source and note: AccessC github.com/alibaba-damo-academy/RynnVLA-Latent · Retrieved 30 October 2026Note. Repository README: “The trained checkpoints are published separately from this code repository, on both Hugging Face and ModelScope.” Only Stage-1 checkpoints are published: “They are not closed-loop control policies”.
- Parameters
- 5.04B Varies by versionSource and note: ParametersA huggingface.co/Alibaba-DAMO-Academy/RynnVLA-Latent-4B · Retrieved 30 October 2026Note. RynnVLA-Latent-4B, model card: “Parameters: 5,042,270,129 (5.042 B)”. RynnVLA-Latent-2B: “Parameters: 2,510,917,041 (2.511 B)”. Technical report, layers of the 4B size: “RynnVLA-Latent-4B 36 2560 4.44 B 768 0.60 B 5.04 B” (language model 4.44 B, action expert 0.60 B, total 5.04 B).
- Base model
- RynnBrainSource and note: Base modelC github.com/alibaba-damo-academy/RynnVLA-Latent · Retrieved 30 October 2026Note. Repository README, Stage-1 training: “initialize from a RynnBrain backbone”. Project page: “RynnBrain VLM + narrow action expert”.
- Licence type
- Free, commercial useSource and note: Licence typeA huggingface.co/Alibaba-DAMO-Academy/RynnVLA-Latent-4B · Retrieved 30 October 2026Note. Weights: “Apache License 2.0.” Code repository: “Apache License 2.0”.
Specifications
11/14 specifications disclosed 3 not stated
1 field hidden 2 not-stated fields with a note stay ·
How to read the figuresShow guideClose
Not statedThe developer publishes nothing on itYesThe developer says it doesVersionsDiffers between versions
Select a source marker for source details and notes.
Release1/1 disclosed
Released25 October 2026
Source and note: ReleasedD arxiv.org/abs/2610.34356 · Retrieved 30 October 2026Note. [A] “Submitted on 25 Oct 2026”
Input and output3/4 disclosed
Model typeVision-language-action (VLA)
Source and note: Model typeA huggingface.co/Alibaba-DAMO-Academy/RynnVLA-Latent-4B · Retrieved 30 October 2026Note. Alibaba's own term: “a Mixture-of-Transformers vision-language-action model”.
InputsCamera Language Robot state
Source and note: InputsA huggingface.co/Alibaba-DAMO-Academy/RynnVLA-Latent-4B · Retrieved 30 October 2026Note. Model card: “Given a language instruction and up to six camera views”. Robot state is added in robot post-training: “Proprioceptive state is introduced only at this stage.”
What it controlsArms and hands
Source and note: What it controlsB arxiv.org/html/2610.34356 · Retrieved 30 October 2026Note. Technical report, after robot post-training: “The policy predicts native action chunks, which are passed to the robot control interface for execution.” The real-robot platforms are dual-arm: “Two dual-arm embodiments used for real-world evaluation”. The report does not say whether the mobile base is driven by the policy.
Degrees of freedom controlledNot stated
Source and note: Degrees of freedom controlledB arxiv.org/html/2610.34356 · Retrieved 30 October 2026Note. The report counts the joints of one test platform, not what the model controls: Marvin-WUJI, “yielding a total of 54 degrees of freedom”.
Size and speed1/2 disclosed
Parameters
5.04B
Versions
- RynnVLA-Latent-2B
- 2,510,917,041
- RynnVLA-Latent-4B
- 5,042,270,129
Source and note: ParametersA huggingface.co/Alibaba-DAMO-Academy/RynnVLA-Latent-4B · Retrieved 30 October 2026Note. RynnVLA-Latent-4B, model card: “Parameters: 5,042,270,129 (5.042 B)”. RynnVLA-Latent-2B: “Parameters: 2,510,917,041 (2.511 B)”. Technical report, layers of the 4B size: “RynnVLA-Latent-4B 36 2560 4.44 B 768 0.60 B 5.04 B” (language model 4.44 B, action expert 0.60 B, total 5.04 B).
Control rateNot stated
Source and note: Control rateB arxiv.org/html/2610.34356 · Retrieved 30 October 2026Note. The report gives no control rate: “inference speed is not an evaluation metric”.
Where it runs and what it is built on1/2 disclosed
On robot or cloudNot stated
Base modelRynnBrain
Source and note: Base modelC github.com/alibaba-damo-academy/RynnVLA-Latent · Retrieved 30 October 2026Note. Repository README, Stage-1 training: “initialize from a RynnBrain backbone”. Project page: “RynnBrain VLM + narrow action expert”.
Openness3/3 disclosed
AccessOpen weights (download)
Source and note: AccessC github.com/alibaba-damo-academy/RynnVLA-Latent · Retrieved 30 October 2026Note. Repository README: “The trained checkpoints are published separately from this code repository, on both Hugging Face and ModelScope.” Only Stage-1 checkpoints are published: “They are not closed-loop control policies”.
Licence documentApache-2.0
Source and note: Licence documentA huggingface.co/Alibaba-DAMO-Academy/RynnVLA-Latent-4B · Retrieved 30 October 2026Note. Model card header: “license: apache-2.0”; licence section: “Apache License 2.0.”
Licence typeFree, commercial use
Source and note: Licence typeA huggingface.co/Alibaba-DAMO-Academy/RynnVLA-Latent-4B · Retrieved 30 October 2026Note. Weights: “Apache License 2.0.” Code repository: “Apache License 2.0”.
The developer’s own claims2/2 disclosed
Learns new tasksNeeds retraining
Source and note: Learns new tasksB arxiv.org/html/2610.34356 · Retrieved 30 October 2026Note. Technical report, robot post-training: “fine-tune the full model on demonstrations from that embodiment”.
Works across robotsYes
Source and note: Works across robotsB arxiv.org/html/2610.34356 · Retrieved 30 October 2026Note. Developer claim: “allowing one interface to support different robots”. Each robot gets its own post-training; the report tests two platforms, Marvin-WUJI and Astribot S1: “The second embodiment utilizes the Astribot S1”.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.
Versions
- RynnVLA-Latent-2B
- RynnVLA-Latent-4B
The developer has released this model in several versions. Each figure applies to the version its source names.
Notes
Alibaba DAMO Academy publishes RynnVLA-Latent in two sizes, 2B and 4B, as Stage-1 pretraining checkpoints.
The released checkpoints are not ready-made robot controllers: a robot needs its own Stage-2 post-training first.
RynnLAM, the model that labels the pretraining videos with latent actions, is published as a separate checkpoint.
Sources
One line per source, not per figure: a source has one retrieval date.
- Ahuggingface.co/Alibaba-DAMO-Academy/RynnVLA-Latent-4BRetrieved 30 October 2026
- Barxiv.org/html/2610.34356Retrieved 30 October 2026
- Cgithub.com/alibaba-damo-academy/RynnVLA-LatentRetrieved 30 October 2026
- Darxiv.org/abs/2610.34356Retrieved 30 October 2026