InternVLA-A1

[A] "Submitted on 5 Jan 2026" ⁠C

Hugging Face model card tag: "vision-language-action-model". ⁠B

Technical report: the multi-view observation "is encoded into visual tokens via the integrated vision encoder, while language instructions" are converted into text tokens. The action expert is conditioned on these features "together with the proprioceptive state". ⁠A

Technical report: "the action expert predicts a target action chunk". The real-robot tests: "These platforms cover diverse bimanual manipulation capabilities". ⁠A

Technical report, table of the 3B release (the main checkpoint InternVLA-A1-3B): "InternVLA-A1 (3B) 3.2B Qwen3-VL (2.13B)". The 2B scale: "InternVLA-A1 (2B) 1.8B InternVL3 (0.94B)". ⁠A

Model inference speed on one GPU, not motor control: "both InternVLA-A1 (2B) and InternVLA-A1 (3B) run at approximately 13 Hz with torch.compile on one RTX 4090 GPU". ⁠A

Technical report: the 3B release uses Qwen3-VL as understanding expert ("InternVLA-A1 (3B) 3.2B Qwen3-VL (2.13B)"); the 2B release uses InternVL3. Abstract: "Building upon InternVL3 and Qwen3-VL, we instantiate InternVLA-A1 at 2B and 3B parameter scales." ⁠A

Hugging Face model card, public repository: "we release the InternVLA-A1 series". ⁠B

Model card header: "license: cc-by-nc-sa-4.0". "All the code within this repo are under CC BY-NC-SA 4.0." ⁠B

Model card header: "license: cc-by-nc-sa-4.0". ⁠B

Developer claim: training ends with "a large-scale pre-training stage on heterogeneous datasets, followed by a task-specific post-training stage". ⁠A

Developer claim: "We benchmark the policies on three physical robot embodiments: Agibot Genie-1, ARX Lift-2, and ARX AC One". ⁠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.