VLA Gallery / All models / Flow·Diffusion VLA

SmolVLA

Hugging Face, Sorbonne University, valeo.ai, ENS Paris-Saclay · 2025-06

VLAFlow·Diffusion VLA0.45Bflow matching

0.45B small VLA pretrained on 481 community SO100 datasets, aimed at single-GPU training and deployment on consumer GPUs and CPUs.

Paper

Category: Flow·Diffusion VLA
Type: VLA
Released: 2025-06
Organization: Hugging Face, Sorbonne University, valeo.ai, ENS Paris-Saclay
Open weights: Yes
Parameters: 0.45B (action expert about 100M); variants 0.24B·0.45B·2.25B (Table 2); HF card states 0.5B

Backbone LLM: SmolLM2 · first 16 layers of SmolVLM2-500M
Vision encoder: SigLIP (SmolVLM2)
Action head: flow matching expert about 100M (10 steps) (flow matching)
Chunk: 50 steps
Control rate: Not disclosed
Architecture (original summary): First 16 layers of SmolVLM-2 (blog: SmolVLM2-500M) (N=L/2) + about 100M flow matching action expert (interleaved cross·self-attention); chunk 50, flow 10 steps, 64 visual tokens per frame; asynchronous inference

Pretraining accelerators: GPU×4 (model not disclosed)
Pretraining duration: Not disclosed
Pretraining steps & batch: 200,000 steps, global batch 256, bf16, cosine lr 1e-4→2.5e-6
GPU/TPU-hours: Not disclosed
Fine-tuning: Paper: simulation 100,000 steps·batch 64, real robot 200,000 steps. LeRobot docs: 1× A100 20k steps about 4 h (batch 64), recording about 50 episodes recommended. Hardware guide: 1× A100 40GB·batch 16·50 episodes 5 epochs about 1~2 h, H100 80GB×4·5000 steps·batch 32 27m 49s, peak VRAM about 10~16 GB (batch 8)
Data type: Real-robot teleoperation (community-collected)
Data scale: 481 datasets (paper)/487 (blog), 22.9K episodes · 10.6M frames (blog: under 30k episodes, about 10 million frames)
Main datasets: LeRobot community datasets (SO100, Hugging Face Hub)
Tasks & evaluation: LIBERO average 87.3%, Meta-World 57.3% (0.45B); SO100 Pick-Place 75%·Stacking 90%·Sorting 70% (average 78.3%); SO101 in-distribution 90%, OOD 50%
Robot: SO100, SO101 (low-cost single arm); simulation LIBERO·Meta-World

Inference device & speed: Targets consumer GPU·CPU deployment; asynchronous inference 9.7 s vs synchronous 13.75 s (about 30% shorter); 2× throughput per blog
Inference memory: Not disclosed
Weights bf16 (calc.): 4.5 GB
Where it runs: Local PC (targeting consumer GPU·CPU)

Reinforcement learning: Not used · Imitation learning (behavior cloning) only
Base control: Fixed · No base control (SO100·SO101 low-cost arms)

Key features: 0.45B small VLA pretrained on 481 community SO100 datasets, aimed at single-GPU training and deployment on consumer GPUs and CPUs. Asynchronous inference cut task completion time by about 30%.
Notes: Parameters (B) is the largest Table 2 variant, 2.25B (main model and released weights are 0.45B). No simulation data used in pretraining; task descriptions auto-generated with Qwen2.5-VL-3B-Instruct. 30k GPU-hours is the whole-project figure.
Verification: Confirmed