AY-ROBOTSAI Teleoperation
Sign Up
MID

TinyVLA

Midea Group, China · September 2024

Midea Group AI Lab and East China Normal University

Compact VLAOpen weights, MITAlso written TinyVLA-S, TinyVLA-B, TinyVLA-H, Tiny-VLA
Parameters
1.3 B
Three variants, all reported in the paper: TinyVLA-S at 422M total with 101M trainable, TinyVLA-B at 740M total with 138M...
GPU memory
2.4 GB
weights at bf16, computed. Fits an 8 GB card
Inference latency
14 ms
Single NVIDIA A6000, time for one action prediction, TinyVLA-H (1.3B).
Weights
MIT
MIT for the GitHub repository liyaxuanliyaxuan/TinyVLA.

What it is

TinyVLA is a family of compact vision language action models from Midea Group's AI Lab and East China Normal University that deliberately skips robot scale pretraining. A small LLaVA style backbone with a Pythia language model (400M, 700M or 1.3B parameters) is fine-tuned directly on roughly 100 demonstrations per task, with a DDPM diffusion decoder producing the actions. On an A6000 the 1.3B variant predicts an action in 14 ms against 292 ms for OpenVLA-7B, and on five real Franka tasks it averaged 94.0 percent against OpenVLA's 68.3 percent. The price of that design is generality: it has no zero-shot behaviour, no released generalist policy checkpoint, and its Meta-World average of 31.6 percent sits far below flow matching policies of similar size.

Architecture

Backbone
Llava-Pythia, a LLaVA style VLM with a Pythia language model, released in 400M, 700M and 1.3B sizes. A PaliGemma based TinyVLA-3B is used only for ablations.
Action head
diffusion policy decoder (DDPM, noise prediction) attached during fine-tuning, conditioned on pooled visual features and proprioceptive state through a 3 layer MLP
Parameters
Three variants, all reported in the paper: TinyVLA-S at 422M total with 101M trainable, TinyVLA-B at 740M total with 138M trainable, and TinyVLA-H at 1.3B total with 143M trainable. Headline results use TinyVLA-H. LoRA is applied to the Q, K and V attention weights, so about 5 percent of the transformer is trainable.
Pretraining data
No robot pretraining at all, which is the paper's central claim. The backbone is a compact VLM pretrained on general vision language data only, and the policy is fine-tuned directly on task demonstrations. The real robot results use 100 trajectories per task, against the 970K Open X-Embodiment trajectories that OpenVLA requires.
Embodiments
Franka Emika (single arm, real), bimanual UR5 (real), Meta-World (simulation)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 1.3 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp324.8 GBTraining master weights, rarely used for inference
bf16 or fp162.4 GBThe usual way these checkpoints are served
int81.2 GBQuantised, expect some loss of precision on fine motions
int40.6 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference latency, reported14 ms on Single NVIDIA A6000, time for one action prediction, TinyVLA-H (1.3B). OpenVLA-7B was measured at 292 ms and OpenVLA-1B at 140 ms on the same setup, which is where the paper's claim of roughly 20x lower inference latency comes from. source

No VRAM figure is published anywhere. All latency measurements were taken on a single NVIDIA A6000. The paper states no training hardware requirement.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • Meta-World easy (28 tasks)
    77.6%
    success rateAuthors' own evaluation. Falls to 21.5 on medium, 11.4 on hard and 15.8 on very hard.source
  • Meta-World average over 50 tasks (easy, medium, hard, very hard)
    31.6%
    success rateTinyVLA-H, authors' own evaluation. Diffusion Policy scored 10.5 in the same table. The SmolVLA paper independently reports the same 31.6 figure for TinyVLA.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Real-world Franka (authors' setup) 5 tasks average, 100 demonstrations per task
    94%
    success rateTinyVLA-H, authors' own hardware. OpenVLA scored 68.3 and a diffusion policy 35.3 in the same setup.source
  • Real-world bimanual UR5 (authors' setup) PlaceBread
    76.7%
    success rateAuthors' own hardware. OpenVLA scored 0 percent on all three bimanual tasks because it was pretrained only on single arm Open X-Embodiment data.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Authors' own evaluation on two platforms. On a single Franka arm across five tasks (PlaceTennis, FlipMug, StackCubes, CloseDrawer, OpenBox) TinyVLA-H averaged 94.0 percent against 68.3 percent for OpenVLA and 35.3 percent for a diffusion policy, using 100 demonstrations per task. On a bimanual UR5 it scored 76.7, 36.7 and 30 percent on three tasks while OpenVLA scored 0 percent on all three. Generalisation was probed across novel objects, unseen positions, appearance changes and background shifts.

Fine tuning it yourself

Training data must be converted to the ACT style h5py layout containing action arrays, language annotations, multi-view images and joint positions. Training starts from one of the released Llava-Pythia backbones (lesjie/Llava-Pythia-400M, 700M or 1.3B) with the diffusion head attached. The paper's real robot results use 100 trajectories per task, so the practical entry cost is roughly 100 demonstrations per skill.

Where it helps, where it does not

Strengths

  • 14 ms per action prediction on an A6000, roughly 20x faster than OpenVLA-7B at 292 ms.
  • Needs no robot scale pretraining, so a working policy can be reached from about 100 demonstrations per task.
  • Handles bimanual hardware, where OpenVLA scores zero because of its single arm pretraining.
  • Three backbone sizes (400M, 700M, 1.3B) let the compute budget be chosen up front.
  • MIT licensed code and openly released backbones.

Limits

  • Meta-World average of 31.6 percent is roughly half what current flow matching policies reach on the same benchmark.
  • Performance collapses outside the easy tier: 77.6 percent easy against 11.4 percent hard.
  • No generalist pretrained policy checkpoint is published. Only the vision language backbones are released, so every task needs its own fine-tune.
  • No zero-shot capability by construction, since robot pretraining is skipped deliberately.
  • No published VRAM figure, no control frequency and no action chunk size.
  • Latest paper revision is v5 from May 2025 and no successor checkpoint has been released under the TinyVLA name.
  • Real robot evidence comes from a single lab's hardware.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.

Download the clientDrive a real arm now