VLA Gallery / All models / New in 2026

TurboVLA

Huazhong University of Science and Technology + Huawei · 2026-07

VLANew in 20260.2BNot disclosed

A 0.2B lightweight VLA reporting 31.2 ms inference latency (32 Hz) and 0.9 GB inference VRAM.

Paper

TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAMarXiv 2607.27205

Overview

Category

New in 2026

Type

VLA

Released

2026-07

Organization

Huazhong University of Science and Technology + Huawei

Open weights

Unverified

Parameters

0.2B

Architecture

Backbone LLM

Unverified

Vision encoder

—

Action head

Unverified (Not disclosed)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

Unverified

Training

Pretraining accelerators

Unverified

Pretraining duration

Unverified

Pretraining steps & batch

80k steps (warm-up 10k), effective batch 256 (stage breakdown unverified)

GPU/TPU-hours

Not disclosed

Fine-tuning

Unverified

Data type

Unverified

Data scale

Unverified

Main datasets

Unverified

Tasks & evaluation

LIBERO average 97.7%

Robot

Unverified

Inference

Inference device & speed

31.2 ms, 32 Hz, VRAM 0.9 GB (paper title refers to RTX 4090)

Inference memory

VRAM 0.9 GB (RTX 4090)

Weights bf16 (calc.)

0.4 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not disclosed · Training method unverified

Base control

Fixed · No base control (LIBERO·RoboTwin 2.0 manipulation benchmarks only)

Notes

Key features

A 0.2B lightweight VLA reporting 31.2 ms inference latency (32 Hz) and 0.9 GB inference VRAM. LIBERO average is 97.7%.

Notes

New facts from verification notes (NEW): read from arXiv 2607.27205 v1 HTML. Announcement month based on arXiv ID (submission date not read).

Verification

Confirmed