Shanghai AI Laboratory · robot AI model
InternW0
China · 2026
Sources retrieved: 2 November 2026 (A, B, C, D)
- Model type
- World action modelSource and note: Model typeC github.com/InternRobotics/InternW0 · Retrieved 2 November 2026Note. GitHub README: “InternW0 is a physical world-action model”.
- What it controls
- Arms and handsSource and note: What it controlsA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card: “a faster ActionDiT expert generates action chunks”. Technical report, pipetting task: “It requires a 20-DoF dexterous hand to achieve stable grasping”.
- Inputs
- Camera Language Robot state TouchSource and note: InputsA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card: “an external frozen DINOv3 ViT-B/16 observation encoder”; “an external frozen Wan2.2 VAE and UMT5 text encoder”; “a 37-dimensional canonical action/proprio interface”. Touch only after fine-tuning: “contact-aware post-training incorporates force and tactile signals for contact-rich manipulation”.
- Access
- Open weights (download)Source and note: AccessC github.com/InternRobotics/InternW0 · Retrieved 2 November 2026Note. GitHub README, public Hugging Face repositories (Base, RoboTwin-clean and LIBERO weights: “Base, RoboTwin-clean, and LIBERO .pt weights”): “All released InternW0 checkpoints are available in the InternW0 Hugging Face collection”.
- Parameters
- 7.35BSource and note: ParametersA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card, count for the released model-only checkpoint (“releases the model-only InternW0-Base initialization”): “Parameters/tensor elements: 7,352,528,643 (serialized in float32)”. The frozen Wan2.2 VAE, text encoder and DINOv3 encoder are external. The video planner is “a Wan2.2-TI2V-5B-based VideoDiT planner (30 layers, hidden size 3072)”.
- Base model
- Wan2.2-TI2V-5BSource and note: Base modelA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card: “a Wan2.2-TI2V-5B-based VideoDiT planner (30 layers, hidden size 3072)”.
- Licence type
- Free, commercial useSource and note: Licence typeA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card header: “license: mit”.
Specifications
12/14 specifications disclosed 2 not stated
0 fields 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 does
Select a source marker for source details and notes.
Release1/1 disclosed
Released23 September 2026
Source and note: ReleasedD arxiv.org/abs/2609.27656 · Retrieved 2 November 2026Note. [A] “Wed, 23 Sep 2026 10:20:56 UTC”
Input and output3/4 disclosed
Model typeWorld action model
Source and note: Model typeC github.com/InternRobotics/InternW0 · Retrieved 2 November 2026Note. GitHub README: “InternW0 is a physical world-action model”.
InputsCamera Language Robot state Touch
Source and note: InputsA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card: “an external frozen DINOv3 ViT-B/16 observation encoder”; “an external frozen Wan2.2 VAE and UMT5 text encoder”; “a 37-dimensional canonical action/proprio interface”. Touch only after fine-tuning: “contact-aware post-training incorporates force and tactile signals for contact-rich manipulation”.
What it controlsArms and hands
Source and note: What it controlsA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card: “a faster ActionDiT expert generates action chunks”. Technical report, pipetting task: “It requires a 20-DoF dexterous hand to achieve stable grasping”.
Degrees of freedom controlledNot stated
Source and note: Degrees of freedom controlledA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. The model card gives a shared interface, not the joints of one robot: “a 37-dimensional canonical action/proprio interface”.
Size and speed2/2 disclosed
Parameters7.35B
Source and note: ParametersA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card, count for the released model-only checkpoint (“releases the model-only InternW0-Base initialization”): “Parameters/tensor elements: 7,352,528,643 (serialized in float32)”. The frozen Wan2.2 VAE, text encoder and DINOv3 encoder are external. The video planner is “a Wan2.2-TI2V-5B-based VideoDiT planner (30 layers, hidden size 3072)”.
Control rate16.47Hz
Source and note: Control rateB internrobotics.github.io/internw0 · Retrieved 2 November 2026Note. Model-side policy updates, not motor control: “corresponding to a maximum model-side update rate of 16.47 Hz”; “60.73 ms per critical-path action update on RTX 5090D”.
Where it runs and what it is built on1/2 disclosed
On robot or cloudNot stated
Source and note: On robot or cloudB internrobotics.github.io/internw0 · Retrieved 2 November 2026Note. The project page names the GPU used for the latency figure, not where the model runs: “60.73 ms per critical-path action update on RTX 5090D”.
Base modelWan2.2-TI2V-5B
Source and note: Base modelA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card: “a Wan2.2-TI2V-5B-based VideoDiT planner (30 layers, hidden size 3072)”.
Openness3/3 disclosed
AccessOpen weights (download)
Source and note: AccessC github.com/InternRobotics/InternW0 · Retrieved 2 November 2026Note. GitHub README, public Hugging Face repositories (Base, RoboTwin-clean and LIBERO weights: “Base, RoboTwin-clean, and LIBERO .pt weights”): “All released InternW0 checkpoints are available in the InternW0 Hugging Face collection”.
Licence documentMIT
Source and note: Licence documentA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card header: “license: mit”; “under the MIT license”.
Licence typeFree, commercial use
Source and note: Licence typeA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Model card header: “license: mit”.
The developer’s own claims2/2 disclosed
Learns new tasksNeeds retraining
Source and note: Learns new tasksA huggingface.co/InternRobotics/InternW0-Base · Retrieved 2 November 2026Note. Developer claim: “It is intended for RoboTwin target-domain posttraining”. “This artifact is a pre-posttraining initialization, not a deployed robot controller”.
Works across robotsYes
Source and note: Works across robotsD arxiv.org/abs/2609.27656 · Retrieved 2 November 2026Note. Developer claim: “Domain-specific interfaces and soft prompts support heterogeneous embodiments”.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
Shanghai AI Laboratory presents InternW0 as the first model of its InternW series.
The InternW0 technical report was submitted to arXiv on 23 September 2026.
Sources
One line per source, not per figure: a source has one retrieval date.
- Ahuggingface.co/InternRobotics/InternW0-BaseRetrieved 2 November 2026
- Binternrobotics.github.io/internw0Main source. Retrieved 2 November 2026
- Cgithub.com/InternRobotics/InternW0Retrieved 2 November 2026
- Darxiv.org/abs/2609.27656Retrieved 2 November 2026