Open Source RoboticsRobots, software & learning

InternVLA-M1

Robotics model

Dual-system spatially guided Vision-Language-Action model built on Qwen2.5-VL-3B-Instruct and a CogACT-style Diffusion Transformer action head with MIT-licensed code, six CC-BY-NC-SA-4.0 (non-commercial) checkpoints, and CUDA, Intel XPU, and CPU runtime support. [1] [2] [3]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Academic and non-commercial robotics researchers exploring Qwen2.5-VL-based VLAs; Teams co-training robot action datasets jointly with multimodal spatial grounding data [1] [3]
Parameters
3,754,622,976 (~3.75B BF16 VLM backbone params in InternRobotics/InternVLA-M1 per HF safetensors metadata [3]
Embodiments
Franka Panda (LIBERO Spatial, Object, Goal, Long), WidowX (SimplerEnv Bridge), Google Robot (SimplerEnv Fractal… [1] [4]
Source rights
Named licenses are documented by part. Includes noncommercial terms. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
Ecosystem
LeRobot

Requirements · Research details and dates

On this page

Overview

InternVLA-M1 (InternRobotics/InternVLA-M1, default branch InternVLA-M1; arXiv:2510.13778, ICLR 2026 *ST4VLA*) is a modular dual-system Vision-Language-Action framework from InternRobotics. It pairs a Qwen/Qwen2.5-VL-3B-Instruct vision-language and spatial-grounding head (chat_with_M1 for image Q&A and bounding-box prediction) with a CogACT-inspired Diffusion Transformer (DiT) action head (predict_action using 10-step DDIM sampling and cfg_scale=1.5) trained under joint dual supervision. [1] [3]

Model card summary

Architecture

Dual-system spatially guided VLA combining a Qwen2.5-VL-3B-Instruct language/spatial-grounding head (Qwen2_5_VLForConditionalGeneration, 36 layers, hidden_size 2048) with a CogACT-inspired Diffusion Transformer (DiT) action head sampled via 10-step DDIM with CFG.

[1] [3]

Base model

Qwen/Qwen2.5-VL-3B-Instruct

[1] [3]

Parameter count

3,754,622,976 (~3.75B BF16 VLM backbone params in InternRobotics/InternVLA-M1 per HF safetensors metadata; full dual-system .pt action checkpoints are 7.79–8.32 GB)

Exact combined VLM + DiT action-head parameter count is not exposed for the single-file .pt checkpoints.

[3]

Input modalities
rgblanguage

Multi-view RGB images (e.g., two views in quickstart demo) and English language instructions or spatial grounding prompts.

[1]

Output modalities
ee-pose-deltagripper-command

Normalized continuous action chunks [B, T, action_dim] for LIBERO and SimplerEnv manipulation tasks (plus text/bounding-box spatial grounding via chat_with_M1).

[1] [4]

Action space

Continuous normalized action chunks [B, T, action_dim] generated by a DiT action head using DDIM sampling (default num_ddim_steps=10, cfg_scale=1.5).

[1]

Trained / supported embodiments
  • Franka Panda (LIBERO Spatial, Object, Goal, Long)
  • WidowX (SimplerEnv Bridge)
  • Google Robot (SimplerEnv Fractal VA/VM)
  • Real-robot setups via examples/real_robot

[1] [4]

Training datasets
  • System-2 multimodal spatial grounding pretraining dataset (referenced in InternRobotics/InternVLA-M1 model card)
  • RT-1 and BridgeData V2 (for InternVLA-M1-Pretrain-RT-1-Bridge)
  • Curated OXE and LIBERO multi-task datasets (IPEC-COMMUNITY / GR00T-LeRobot format)

[1] [3]

Interfaces
  • Python InternVLA_M1 API (from_pretrained, chat_with_M1, predict_action)
  • Unified inference server for SimplerEnv and LIBERO
  • GR00T-LeRobot standardized action data loader

[1] [4]

Checkpoints

CheckpointFormatSizeAccessPinned revision
InternRobotics/InternVLA-M1safetensors6.99 GBopence8a34244d1d
InternRobotics/InternVLA-M1-Pretrain-RT-1-Bridgepytorch7.79 GBopen6272240f1c14
InternRobotics/InternVLA-M1-LIBERO-Longpytorch8.32 GBopen67a325f3f01d
InternRobotics/InternVLA-M1-LIBERO-Goalpytorch8.32 GBopenf02a857ef543
InternRobotics/InternVLA-M1-LIBERO-Spatialpytorch8.32 GBopenb1bd418dcc3d
InternRobotics/InternVLA-M1-LIBERO-Objectpytorch8.32 GBopen436c7412d3c1

All 6 checkpoints are ungated on Hugging Face but licensed CC-BY-NC-SA-4.0 (non-commercial).

[1] [3]

Is this for you?

Use cases
  • Dual-supervision spatial Q&A / bounding-box grounding (chat_with_M1) and multi-view robot action prediction (predict_action)
  • Benchmark evaluation on LIBERO (Spatial, Object, Goal, Long) and SimplerEnv (WidowX, Google Robot)
  • Co-training action datasets with multimodal vision-language supervision
  • Running VLA inference on NVIDIA CUDA, Intel GPUs (XPU), or CPU

[1] [4]

Known limitations
  • All six published Hugging Face checkpoints are licensed CC-BY-NC-SA-4.0 (non-commercial, share-alike), even though the GitHub codebase is MIT-licensed
  • Pretrained RT-1-Bridge and all four LIBERO checkpoints are distributed as single PyTorch .pt files (7.79–8.32 GB) rather than safetensors
  • FlashAttention 2 is CUDA-only (Intel XPU and CPU fall back to PyTorch native attention, with minor float32 flow-matching rounding drift noted upstream)
  • Minimum inference VRAM, fine-tuning VRAM, and closed-loop control frequency (Hz) are not specified in the README

[1] [3]

Safety and limitations
  • All six released checkpoints are licensed CC-BY-NC-SA-4.0, prohibiting commercial use of the published weights despite the MIT-licensed codebase.
  • Five of six checkpoints are single-file PyTorch .pt archives requiring trust before deserialization.
  • No robot-side safety guardrails or collision interlocks are bundled in the inference server.

[1] [3]

Compute & fine-tuning

Inference compute
  • Documented hardware: NVIDIA A100, RTX 4090 (LIBERO eval verified); Intel XPU and CPU also supported
  • Precision: bf16
  • Edge / consumer capable: No

Exact minimum VRAM and inference latency_ms are not documented in README.

[1] [4]

Fine-tuning
  • Supported: Yes
  • Methods: Spatial-prompt and action fine-tuning (~2.5 epochs reported for convergence); Multimodal multi-task co-training (examples/CoTraininig/README.md)
  • Recipe: Upstream fine-tuning documentation

[1]

Host computer

Runs on NVIDIA CUDA (LIBERO evaluation verified on A100 and RTX 4090), Intel GPUs (XPU), or CPU with automatic device selection; exact minimum VRAM and latency_ms are not specified.

[1] [4]

Operating system
  • Linux

Standard Linux conda, CUDA/XPU, and flash-attn workflow documented.

[1] · Editorial assessment

Dependencies
  • Python 3.10
  • PyTorch (CUDA wheels or Intel XPU wheels)
  • transformers (4.50.3 recorded in config.json)
  • flash-attn (CUDA only; optional on XPU/CPU)
  • Qwen/Qwen2.5-VL-3B-Instruct (base VLM)

[1] [3]

Skills
  • Python 3.10 and conda
  • PyTorch (CUDA, Intel XPU, or CPU)
  • GR00T-LeRobot / OXE / LIBERO dataset configuration

Inferred from README environment setup, LIBERO service workflow, and acknowledgements.

[1] [4] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [1] · Model card / documentation [1]

  • Clone InternRobotics/InternVLA-M1 (default branch InternVLA-M1) and create a Python 3.10 conda environment
  • Install requirements.txt, install flash-attn (skip on Intel XPU/CPU), and run pip install -e .
  • Download a checkpoint from Hugging Face (e.g., InternRobotics/InternVLA-M1 or a LIBERO/RT-1-Bridge .pt checkpoint)
  • Load via InternVLA_M1.from_pretrained(saved_model_path) and call chat_with_M1 or predict_action(batch_images, instructions, cfg_scale=1.5, use_ddim=True, num_ddim_steps=10)

[1]

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
LIBERO (Spatial, Objects, Goal, Long)95.9% average (Spatial 98.0%, Objects 99.0%, Goal 93.8%, Long 92.6%)Upstream-reported results in README and examples/LIBERO/README.md (verified on A100 and RTX 4090); not independently reproduced by OSR.
SimplerEnv (WidowX, Google Robot VA & VM)WidowX 71.7%, Google Robot VA 76.0%, Google Robot VM 80.7%Upstream-reported results in README table (arXiv:2510.13778 / ICLR 2026 ST4VLA); not independently reproduced by OSR.

[1] [4]

Project notes

Six checkpoints, hardware portability, and data loader reuse

Six ungated checkpoints are hosted on Hugging Face: the base InternRobotics/InternVLA-M1 system-2 model (3,754,622,976 BF16 parameters, 6.99 GB safetensors + .bin), InternVLA-M1-Pretrain-RT-1-Bridge (7.79 GB .pt), and four LIBERO task-suite checkpoints (LIBERO-Long, LIBERO-Goal, LIBERO-Spatial, LIBERO-Object, 8.32 GB .pt each). Notably for hardware flexibility, deployment and evaluation scripts automatically select NVIDIA CUDA, Intel GPUs (XPU), or CPU without code changes; Intel XPU skips CUDA-only flash-attn and uses PyTorch's native attention (with minor float32 rounding drift noted upstream). Data loading reuses the GR00T-LeRobot format alongside IPEC-COMMUNITY OXE/LIBERO mixtures, and the repository includes a unified inference server for SimplerEnv and LIBERO plus multimodal co-training recipes. [1] [3] [4]

Upstream-reported benchmarks and undocumented compute limits

In the repository README and examples/LIBERO/README.md, InternVLA-M1 reports a 95.9% average across LIBERO suites (Spatial 98.0%, Objects 99.0%, Goal 93.8%, Long 92.6%) and SimplerEnv scores of 71.7% on WidowX, 76.0% on Google Robot Variant Aggregation, and 80.7% on Google Robot Visual Matching. Separately, upstream reports that the LIBERO evaluation workflow runs on NVIDIA A100 and RTX 4090 GPUs. However, upstream does not specify minimum inference or fine-tuning VRAM, per-step inference latency, or closed-loop control frequency (Hz), and the exact combined VLM + DiT parameter count is not exposed in Hugging Face metadata for the five single-file .pt checkpoints. All benchmark figures are upstream-reported; Open Source Robotics has not run training, fine-tuning, inference, or robot evaluations for this profile. [1] [3] [4]

What we checked

None. Prepared from public upstream sources and Hugging Face API metadata without hands-on training, fine-tuning, or inference.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
Oct 9, 2026 [1]
Latest release
Nov 24, 2025 [3]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
InternRobotics/InternVLA-M1@7f42cf279a3b15d79507b632d37921f5f2f073ba (branch InternVLA-M1); HF InternRobotics/InternVLA-M1@ce8a34244d1da9cf5136743f28f1a619168a5185 · combined hardware and software repository

Separate license matrix

Active dual-system Qwen2.5-VL + DiT VLA framework with MIT-licensed code and six ungated checkpoints; classified as partial openness because all published model weights carry a non-commercial CC-BY-NC-SA-4.0 license.

LayerAvailability / license
Codeavailable · MIT
Verified MIT License (InternVLA Team) in root LICENSE file. [1] [2]
Model weightsavailable · CC-BY-NC-SA-4.0 (non-commercial, share-alike)
All 6 InternRobotics/InternVLA-M1* Hugging Face checkpoints declare license: cc-by-nc-sa-4.0. [3]
Training datasetsavailable · Mixed (public IPEC-COMMUNITY OXE/LIBERO datasets have per-dataset licenses; system-2 spatial grounding pretraining dataset is referenced in model card without a standalone license link)
Partial availability: public IPEC-COMMUNITY OXE/LIBERO datasets are available, while the system-2 spatial grounding pretraining dataset is not linked as a standalone public dataset. [1] [3]
Documentationavailable · MIT
[1] [2]

Critical code-vs-weights licensing split

Although the InternRobotics/InternVLA-M1 codebase and documentation are released under the MIT License (verified in the root LICENSE file), all six published Hugging Face model checkpoints declare CC-BY-NC-SA-4.0, which prohibits commercial use and requires share-alike terms on derivative weights. Commercial teams can reuse the MIT training framework on their own data and base models (subject to Qwen/Qwen2.5-VL-3B-Instruct and dataset licenses), but cannot deploy the published InternRobotics/InternVLA-M1* checkpoints commercially. [1] [2] [3]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Initial 0.8.0-draft.1 model profile prepared from InternRobotics/InternVLA-M1@7f42cf279a3b and six HF checkpoints.