S-Lab, Nanyang Technological University · robot AI model
NEO-VLA
Singapore · 2026
Sources retrieved: 9 November 2026 (A, B, C, D)
- Model type
- Vision-language-action (VLA)Source and note: Model typeD arxiv.org/abs/2611.06918 · Retrieved 9 November 2026Note. Developer wording: “We introduce NEO-VLA, a structurally encoder-free vision-language-action framework.”
- What it controls
- Arms and handsSource and note: What it controlsA github.com/EvolvingLMMs-Lab/NEO/tree/main/VLAKit · Retrieved 9 November 2026Note. README: “The action expert predicts 30-step chunks for two arms and a body slot.” It adds: “The default action representation supports 26 dimensions per arm and 8 body dimensions”.
- Inputs
- Camera Language Robot stateSource and note: InputsA github.com/EvolvingLMMs-Lab/NEO/tree/main/VLAKit · Retrieved 9 November 2026Note. README: “Multi-view observations, camera geometry, instructions and robot state share one prefix.”
- Access
- Open weights (download)Source and note: AccessC huggingface.co/weepiess2383/NEO-VLA · Retrieved 9 November 2026Note. Model card: “Download checkpoints with” the huggingface_hub library. Hugging Face lists the repository as public and not gated.
- Base model
- NEO-2BSource and note: Base modelA github.com/EvolvingLMMs-Lab/NEO/tree/main/VLAKit · Retrieved 9 November 2026Note. README: “The 40-layer NEO-2B backbone processes previous and current observations, text and state.”
- Licence document
- Apache-2.0Source and note: Licence documentC huggingface.co/weepiess2383/NEO-VLA · Retrieved 9 November 2026Note. The model card header says “license: apache-2.0”. Its text adds: “The code is distributed under Apache-2.0. Backbone weights and source datasets retain their upstream terms; the code license does not relicense those assets.”
Specifications
8 of 14 stated 6 not stated
0 fields hidden 6 not-stated fields with a note stay ·
How to read the figuresShow guideClose
- not statedNo source
- 0Reported as zero
- ASource
Select a source marker for source details and notes.
Release1/1 disclosed
Released4 November 2026
Source and note: ReleasedD arxiv.org/abs/2611.06918 · Retrieved 9 November 2026Note. [A] arXiv submission history: “[v1] Wed, 4 Nov 2026 18:59:58 UTC”.
Input and output3/4 disclosed
Model typeVision-language-action (VLA)
Source and note: Model typeD arxiv.org/abs/2611.06918 · Retrieved 9 November 2026Note. Developer wording: “We introduce NEO-VLA, a structurally encoder-free vision-language-action framework.”
InputsCamera Language Robot state
Source and note: InputsA github.com/EvolvingLMMs-Lab/NEO/tree/main/VLAKit · Retrieved 9 November 2026Note. README: “Multi-view observations, camera geometry, instructions and robot state share one prefix.”
What it controlsArms and hands
Source and note: What it controlsA github.com/EvolvingLMMs-Lab/NEO/tree/main/VLAKit · Retrieved 9 November 2026Note. README: “The action expert predicts 30-step chunks for two arms and a body slot.” It adds: “The default action representation supports 26 dimensions per arm and 8 body dimensions”.
Degrees of freedom controlledNot stated
Source and note: Degrees of freedom controlledA github.com/EvolvingLMMs-Lab/NEO/tree/main/VLAKit · Retrieved 9 November 2026Note. The README gives action dimensions without naming a robot: “The default action representation supports 26 dimensions per arm and 8 body dimensions”.
Size and speed0/2 disclosed
ParametersNot stated
Source and note: ParametersB arxiv.org/html/2611.06918v2 · Retrieved 9 November 2026Note. The paper gives a figure for the backbone only: “we train the 40-layer encoder-free vision-language backbone with 2B parameters”.
Control rateNot stated
Source and note: Control rateA github.com/EvolvingLMMs-Lab/NEO/tree/main/VLAKit · Retrieved 9 November 2026Note. The model predicts action chunks: “The action expert predicts 30-step chunks for two arms and a body slot.”
Where it runs and what it is built on1/2 disclosed
On robot or cloudNot stated
Source and note: On robot or cloudA github.com/EvolvingLMMs-Lab/NEO/tree/main/VLAKit · Retrieved 9 November 2026Note. README: “neo_vla.serve turns a post-trained step into an HTTP policy service for a real robot.”
Base modelNEO-2B
Source and note: Base modelA github.com/EvolvingLMMs-Lab/NEO/tree/main/VLAKit · Retrieved 9 November 2026Note. README: “The 40-layer NEO-2B backbone processes previous and current observations, text and state.”
Openness2/3 disclosed
AccessOpen weights (download)
Source and note: AccessC huggingface.co/weepiess2383/NEO-VLA · Retrieved 9 November 2026Note. Model card: “Download checkpoints with” the huggingface_hub library. Hugging Face lists the repository as public and not gated.
Licence documentApache-2.0
Source and note: Licence documentC huggingface.co/weepiess2383/NEO-VLA · Retrieved 9 November 2026Note. The model card header says “license: apache-2.0”. Its text adds: “The code is distributed under Apache-2.0. Backbone weights and source datasets retain their upstream terms; the code license does not relicense those assets.”
Licence typeNot stated
Source and note: Licence typeC huggingface.co/weepiess2383/NEO-VLA · Retrieved 9 November 2026Note. The model card header gives Apache-2.0, but its text says backbone weights and source datasets keep their upstream terms: “The code is distributed under Apache-2.0. Backbone weights and source datasets retain their upstream terms; the code license does not relicense those assets.”
The developer’s own claims1/2 disclosed
Learns new tasksNeeds retraining
Source and note: Learns new tasksB arxiv.org/html/2611.06918v2 · Retrieved 9 November 2026Note. Developer claim. Paper: real-world tasks use fine-tuning, “models are fine-tuned on only 50 demonstration episodes”. Model card: “These checkpoints support robot-learning research and require task-specific adaptation and evaluation before use on a physical robot.”
Works across robotsNot stated
Source and note: Works across robotsB arxiv.org/html/2611.06918v2 · Retrieved 9 November 2026Note. The paper describes the interface, not robots it runs on: “Components absent from a particular embodiment are simply masked, allowing the same token interface to accommodate different robot configurations.”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.
Notes
The NEO-VLA code is published by the EvolvingLMMs-Lab GitHub organisation, and the paper lists three affiliations: S-Lab (Nanyang Technological University), SenseTime Research and Ropedia. The weights are hosted on the Hugging Face account weepiess2383.
The model card says the checkpoints are for robot-learning research and need task-specific adaptation and evaluation before use on a physical robot.
The model card says the released backbone was fully fine-tuned during joint vision-language and action training, so it differs from the original NEO-2B checkpoint.
Sources
One line per source, not per figure: a source has one retrieval date.
- Agithub.com/EvolvingLMMs-Lab/NEO/tree/main/VLAKitMain source. Retrieved 9 November 2026
- Barxiv.org/html/2611.06918v2Retrieved 9 November 2026
- Chuggingface.co/weepiess2383/NEO-VLARetrieved 9 November 2026
- Darxiv.org/abs/2611.06918Retrieved 9 November 2026
