Skip to content
embodied.guide

Back to the AI models

Shanghai AI LaboratoryChina2025

InternVLA-M1

An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.

Robotsnot stated

Key figures as stated by the developer

6 of 6 stated
  • Model type ⁠B vision-language-action
  • Controls ⁠A arms and hands
  • Inputs ⁠A camera language robot state
  • Access ⁠B open download
  • Parameters ⁠A 4,100,000,000
  • Base model ⁠C Qwen2.5-VL-3B-Instruct
How to read the figures
14kgA30.9lbA
The letter in the grey square after a figure points to the source the figure came from. Robots with several sources carry several letters.
74.5lb
Imperial figures are our own conversion unless the manufacturer states an imperial figure itself. The source letter always belongs to the metric figure — a conversion has no source of its own.

The four states

Versions and notes in the key figures (6)

Model type

Developer wording: "is a spatially guided vision-language-action framework for generalist robot policy.".

Controls

The real-world test uses a single arm: "we employ a Franka Research 3 robotic arm equipped with a Robotiq 2F-85 gripper".

Inputs

On the LIBERO benchmark "the policy takes as input both wrist-mounted and third-person camera views". The README example passes a text instruction ("Pick up the apple and place it on the plate."). The action expert includes "a lightweight state encoder (0.4 M)".

Access

The GitHub roadmap lists "0918: Release model weights"; the checkpoints are on Hugging Face.

Parameters

"In total, InternVLA-M1 comprises approximately 4.1B parameters."

Base model

The model card lists "- Qwen/Qwen2.5-VL-3B-Instruct" under base_model.

Full specifications

8 of 12 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
Input and output
Model typevision-language-actionB

Developer wording: "is a spatially guided vision-language-action framework for generalist robot policy.".

Inputscamera language robot stateA

On the LIBERO benchmark "the policy takes as input both wrist-mounted and third-person camera views". The README example passes a text instruction ("Pick up the apple and place it on the plate."). The action expert includes "a lightweight state encoder (0.4 M)".

Controlsarms and handsA

The real-world test uses a single arm: "we employ a Franka Research 3 robotic arm equipped with a Robotiq 2F-85 gripper".

Degrees of freedom controllednot stated
Size and speed
Parameters4,100,000,000A

"In total, InternVLA-M1 comprises approximately 4.1B parameters."

Control ratenot stated

Note The paper reports inference speed, not a control rate: "With FlashAttention, the VLM component achieves inference speeds of approximately 10 FPS." ⁠A
Where it runs and what it is built on
Runs wherenot stated

Note The paper says: "During inference, the system runs on a single RTX 4090 GPU with around 12 GB of memory usage." ⁠A

Base modelQwen2.5-VL-3B-InstructC

The model card lists "- Qwen/Qwen2.5-VL-3B-Instruct" under base_model.
Openness
Accessopen downloadB

The GitHub roadmap lists "0918: Release model weights"; the checkpoints are on Hugging Face.

LicenseCC BY-NC-SA 4.0C

The model card on Hugging Face says "license: cc-by-nc-sa-4.0". The GitHub repository (code) is under the MIT licence (""spdx_id": "MIT"").
The developer’s own claims
Learns new tasksneeds retrainingA

Developer claim: "our model exhibits strong generalization to unseen objects and instructions under few-demonstration fine-tuning".

Works across robotsnot stated

Note The paper builds on "an embodiment-agnostic spatial prior"; for the benchmarks, "We fine-tune the model on each suite independently". ⁠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.

Notes

  • The developers call InternVLA-M1 a spatially guided vision-language-action framework for generalist robot policy (GitHub README: "is a spatially guided vision-language-action framework for generalist robot policy.").
  • The project page is signed Intern Robotics, Shanghai AI Laboratory.
  • The technical report was posted to arXiv on 15 October 2025.

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.