Open Source RoboticsRobots, software & learning

SmolVLA

Robotics model

Compact 450M-parameter flow-matching Vision-Language-Action base model built on SmolVLM2-500M-Video-Instruct for affordable single-GPU fine-tuning and consumer hardware deployment via LeRobot, with Apache-2.0 weights and pretraining reproduction kit (VLAb). [1] [2] [3]

Setup & usage →CheckpointsModel hub / homepageSource codeGetting started

For
Low-cost robot builders (SO-100/SO-101, LeRobot users); Robotics researchers; Single-GPU VLA practitioners [1] [3]
Parameters
450,046,176 (~450M [1]
Embodiments
so-101, SO-100, LeRobot community-compatible arms (padded up to 32-D state/action), … [1] [2] [3]
Source rights
Named licenses are documented by part. 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

SmolVLA (lerobot/smolvla_base) is a 450,046,176-parameter (~450M) flow-matching Vision-Language-Action base model built on HuggingFaceTB/SmolVLM2-500M-Video-Instruct. It is packaged for consumer-hardware fine-tuning and deployment through lerobot[smolvla], with a separate multi-GPU pretraining reproduction repository (huggingface/vlab). Note that checkpoints produced by huggingface/vlab may be incompatible with newer LeRobot normalization formats and may require the documented normalization migration before LeRobot fine-tuning or inference; this caveat applies specifically to VLAb-produced checkpoints. [1] [2] [3]

Model card summary

Architecture

SmolVLM2-500M-Video-Instruct vision-language backbone paired with a compact flow-matching action expert (10 denoising steps by default).

[1] [2]

Base model

HuggingFaceTB/SmolVLM2-500M-Video-Instruct

[1]

Parameter count

450,046,176 (~450M; 446,772,624 BF16 + 3,273,552 F32 per safetensors metadata)

Exact tensor count verified via Hugging Face API safetensors metadata.

[1]

Input modalities
rgbproprioceptionlanguage

Up to 3 RGB cameras (resized with padding to 512x512; default input shape 3x256x256), proprioceptive state (default 6-D, padded up to max_state_dim: 32), optional English language instruction.

[1] [2]

Output modalities
joint-positionsgripper-command

Continuous action chunks (default shape 6, padded up to max_action_dim: 32) for joint/gripper control.

[1] [2]

Action space

Continuous flow-matching action chunks with chunk_size: 50, n_action_steps: 50, default 6-DoF action vector padded up to max_action_dim: 32, and 10 flow-matching integration steps (num_steps: 10).

Verified directly from config.json at revision 5e8d12a6e297.

[2]

Trained / supported embodiments
  • so-101
  • SO-100
  • LeRobot community-compatible arms (padded up to 32-D state/action)
  • LIBERO simulation (Franka)

Quickstart demonstrates lerobot/libero; designed for affordable LeRobot arms.

[1] [2] [3]

Training datasets
  • LeRobot community datasets on Hugging Face Hub (multi-dataset mixture documented in huggingface/vlab)

Trained on community-driven Hugging Face robotics datasets; exact dataset list is maintained in VLAb reproduction configs.

[1] [3]

Interfaces
  • LeRobot Python API (SmolVLAPolicy)
  • Hugging Face Hub safetensors
  • Accelerate CLI (VLAb)

[1] [2] [3]

Checkpoints

CheckpointFormatSizeAccessPinned revision
lerobot/smolvla_basesafetensors0.85 GBopen5e8d12a6e297

Ungated safetensors checkpoint (0.85 GB) verified via Hugging Face API.

[1]

Is this for you?

Use cases
  • Fine-tuning compact VLA policies on custom LeRobot datasets
  • Consumer-GPU robot manipulation inference
  • Multi-GPU VLA pretraining reproduction via VLAb

[1] [3]

Known limitations
  • Intended as a base model for task-specific fine-tuning rather than zero-shot deployment on arbitrary new setups
  • Runtime inference and fine-tuning depend on the separate huggingface/lerobot package

[1] [3]

Safety and limitations
  • Base checkpoint requires fine-tuning on target robot/task demonstrations before real-robot deployment.
  • Trained on heterogeneous community-contributed LeRobot datasets whose quality and licensing vary per upstream dataset.

[1] [3]

Compute & fine-tuning

Inference compute
  • Documented hardware: Consumer single GPU / CPU (per model card)
  • Precision: bf16
  • Edge / consumer capable: Yes

0.85 GB BF16/F32 safetensors footprint supports consumer hardware and CPU/GPU inference; exact minimum VRAM and latency_ms not stated in model card.

[1] · Editorial assessment

Fine-tuning

Pinned config.json (S002) sets freeze_vision_encoder: true, train_expert_only: true, and train_state_proj: true by default; VLAb-produced checkpoints may require normalization migration before LeRobot fine-tuning or inference (S003).

[1] [3]

Host computer

Designed to be trainable on a single GPU and deployable on consumer hardware (CUDA or CPU supported in quickstart).

[1]

Operating system
  • Linux

Standard Python/PyTorch/conda Linux workflow documented.

[1] [3] · Editorial assessment

Dependencies
  • Python
  • PyTorch
  • lerobot[smolvla]
  • HuggingFaceTB/SmolVLM2-500M-Video-Instruct
  • Accelerate (for VLAb pretraining)

[1] [3]

Skills
  • Python
  • PyTorch
  • LeRobot CLI/API

Inferred from documented Python/PyTorch/CLI setup and training dependencies.

[1] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [1]

  • Install pip install 'lerobot[smolvla]'
  • Load lerobot/smolvla_base via SmolVLAPolicy.from_pretrained
  • Run make_pre_post_processors and policy.select_action or fine-tune on a LeRobotDataset
  • For pretraining reproduction from scratch, clone huggingface/vlab and create conda env

[1] [3]

Upstream-reported benchmarks

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

BenchmarkUpstream-reported scoreScope & caveats
SmolVLA paper evaluations (arXiv:2506.01844; LIBERO, Meta-World, real-robot tasks)See arXiv:2506.01844 (model card states performance matching much larger VLAs)Upstream-reported benchmark claims; exact tables in arXiv:2506.01844 not independently verified by OSR.

[1]

Project notes

Compact architecture and padded 32-D action space

At the pinned Hub revision (5e8d12a6e297), config.json specifies continuous flow-matching action chunks with chunk_size: 50, n_action_steps: 50, and num_steps: 10 denoising integration steps. Inputs accept up to three RGB camera streams (resized with padding to 512x512), proprioceptive state (default 6-D, padded up to max_state_dim: 32), and optional English language instructions, outputting continuous joint/gripper actions padded up to max_action_dim: 32. For fine-tuning, config.json sets freeze_vision_encoder: true, train_expert_only: true, and train_state_proj: true by default—meaning the default configuration updates the action expert and state projection rather than every parameter in the model. The checkpoint is intended as a base model for task-specific fine-tuning on target robot demonstrations rather than zero-shot control on arbitrary hardware. [1] [2]

Deployment footprint and upstream-reported benchmarks

The ungated safetensors checkpoint totals 0.85 GB (BF16/F32), and the model card documents single-GPU training as well as consumer GPU or CPU inference via LeRobot. Exact minimum VRAM and closed-loop control frequency are not pinned as single scalars in config.json and depend on the host and asynchronous inference setup. Benchmark claims in arXiv:2506.01844 (covering LIBERO, Meta-World, and real-robot tasks) are upstream-reported; Open Source Robotics has not run training, fine-tuning, inference, or robot evaluations for this profile. [1] [2]

What we checked

None. Prepared from public upstream sources without hands-on hardware execution.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
Oct 6, 2026 [1]
Latest release
Oct 6, 2026 [1]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
May 21, 2027
Reviewed revision
HF lerobot/smolvla_base@5e8d12a6e2975b0e5e5fce7c8caf47c371d257b6; huggingface/vlab@10558f6f958902c1b5ff5eed76ff5766fab6f64b · combined hardware and software repository

Separate license matrix

Standalone 450M open-weights flow-matching VLA model with verified Apache-2.0 weights and Apache-2.0 pretraining reproduction kit (VLAb); profiled narrowly as a model artifact while linking to existing lerobot runtime.

LayerAvailability / license
Codeavailable · Apache-2.0
GitHub API reports NOASSERTION due to formatting, but root LICENSE file at commit 10558f6f9589 is verbatim Apache License Version 2.0. [3] [4]
Model weightsavailable · Apache-2.0
Model card YAML declares license: apache-2.0; base model SmolVLM2-500M-Video-Instruct is also Apache-2.0. [1]
Training datasetsavailable · Mixed (community-contributed Hugging Face LeRobot datasets; per-dataset licenses apply)
Pretraining relies on multiple public community datasets on Hugging Face Hub with individual dataset licenses. [1] [3]
Documentationavailable · Apache-2.0
[1]

Separate code, weights, and community dataset rights

Both the lerobot/smolvla_base checkpoint and the huggingface/vlab pretraining repository carry Apache-2.0 licenses (verified directly in vlab's root LICENSE file and the model card YAML, alongside the Apache-2.0 SmolVLM2-500M-Video-Instruct base model). Pretraining draws on community-contributed LeRobot datasets hosted on Hugging Face Hub, where individual upstream dataset licenses and data quality vary by contributor and must be reviewed separately before commercial reuse. [1] [3] [4]

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 draft prepared from HF revision 5e8d12a6e297 and vlab commit 10558f6f9589.