TurboVLA: A 0.2B Vision-Language-Action Model Hitting 32 Hz with <1 GB VRAM — No LLM Required

A 0.2B-parameter model achieves 31.2 ms inference latency, 0.9 GB VRAM usage, and a 97.7% average success rate on LIBERO — all on a consumer RTX 4090 — while discarding the assumption that a VLA model must have an LLM as its central hub.

Key numbers

| Metric | TurboVLA | Comparison | |---|---|---| | Parameters | 0.2B | OpenVLA 7B, RT-2 55B | | Inference VRAM | 0.9 GB | OpenVLA ~16 GB | | Inference latency | 31.2 ms (32 Hz) | OpenVLA ~150-300 ms (3-6 Hz) | | LIBERO average success | 97.7% | OpenVLA ~45%, RT-2 ~57%, π0 65-70% |

Paper info

• Title: *TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM*
• arXiv: https://arxiv.org/abs/2607.27205
• HuggingFace: https://huggingface.co/H-EmbodVis/TurboVLA
• Authors: H Xie et al., H-EmbodVis team
• Published: July 2026

What TurboVLA does

The mainstream VLA paradigm is V→L→A:

• Vision encoder translates images into language tokens
• An LLM fuses visual tokens with the language instruction and generates action tokens
• An action decoder outputs joint angles / end-effector poses

The problem: every inference wakes up a 7B+ LLM, even when the instruction — "put the red block in the blue bowl" — already specifies the skill.

TurboVLA instead uses a V+L→A direct mapping:

• Vision encoder + lightweight language encoder (BERT-scale)
• Bidirectional cross-modal interaction layers (vision and language tokens attend to each other)
• Direct action output

Key observation from the paper: when the instruction already specifies the skill, the execution policy does not need open-ended language generation — the instruction only needs to determine how visual evidence guides action. On the execution path, the LLM's role is redundant.

Three counterintuitive takeaways

1. The 97.7% LIBERO score is not a "small model, easy task" win

LIBERO's Spatial / Object / Goal / Long subsets test spatial understanding, object attributes, goal orientation, and long-horizon tasks. OpenVLA-OFT and RDT-1B score only 23-29% on some subsets. A 97.7% average means TurboVLA is near-perfect on all four — a small model achieving top performance on complex tasks.

2. 32 Hz is the hard threshold for real closed-loop control

In robot control, stable closed-loop feedback generally requires at least 20-30 Hz. OpenVLA's 3-6 Hz only supports open-loop execution with periodic replanning. At 32 Hz, TurboVLA can adjust actions every 31 ms based on fresh visual feedback — a qualitative shift from "runs" to "runs in real time."

3. <1 GB VRAM changes deployment economics

OpenVLA's ~16 GB footprint means one GPU runs essentially one instance. TurboVLA's 0.9 GB means a single RTX 4090 can run 20+ instances in parallel, or co-host the VLA alongside visual SLAM, motion planning, and safety monitoring. It also opens edge deployment: a Jetson Orin Nano (8 GB unified memory) can run it — the robot doesn't need to carry a 4090.

Observations

1. "De-LLM-ification" is a hidden theme in current AI system design

We've spent two years accepting "LLM as universal hub" — VLAs, agents, RAG. TurboVLA shows that when the task boundary is clear and the instruction already specifies the skill, the LLM's open-ended generation is redundant overhead. This parallels other work moving the LLM from "present everywhere" to "present only when needed."

2. Boundary conditions for the V+L→A mapping

TurboVLA's claim has an important premise: the instruction already specifies the skill. LIBERO's instructions are relatively structured. In scenarios with ambiguous instructions requiring open-ended reasoning ("tidy up the desk," "something seems off, take a look"), the V→L→A paradigm may still be necessary. More precisely: TurboVLA delineates the applicability boundary of direct V+L→A mapping — clear instructions, closed skill set, short horizon. Beyond that boundary, the LLM hub retains irreplaceable value. It's not "LLMs are useless" — it's "the LLM's battlefield has shrunk."

3. Bidirectional cross-modal interaction vs. one-way translation

V→L→A implicitly assumes vision must first be "translated" into language tokens — a lossy step that drops visual continuity, spatial relations, and motion information. TurboVLA's bidirectional cross-modal attention skips the translation: visual information fuses with the language instruction directly in a multimodal representation space, and drives continuous actions at execution level.

4. The broader thread

The mainline is getting clearer: AI system design is shifting from "LLM always present" to "LLM on demand" — visual perception, action execution, and constraint checking are all being de-LLM-ified, with the LLM retreating to a position where it is only invoked for open-ended reasoning.

A caveat worth heeding

LIBERO is a simulation benchmark. The sim2real gap is a chronic robotics problem — 97.7% in simulation might drop to 60-70% on real hardware, and the paper does not report detailed sim2real transfer results. Also, 32 Hz closed-loop control is easy in simulation (simulators run at hundreds of Hz); on a real robot, the perception pipeline alone (camera exposure + ISP + USB transfer + preprocessing) can consume 10-15 ms, leaving a much tighter budget than 31.2 ms. End-to-end latency and success rate on real hardware are the final judgment — don't fixate on 97.7% and 32 Hz alone.

Open question

TurboVLA's bidirectional cross-modal interaction layer is essentially a lightweight attention bridge. How compatible is it with LoRA fine-tuning? Could one fine-tune a TurboVLA-style lightweight action head — ideally via ES-style methods — onto an existing OpenVLA model, reusing its vision representations? If so, this "VLA distillation" path might reach deployment faster than training TurboVLA from scratch.