X-VLA is a Vision-Language-Action foundation model that uses soft prompts to handle cross-embodiment and cross-domain robot control within a unified Transformer architecture. Original paper: X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model For full installation details (including optional video dependencies such as ffmpeg for torchcodec), see the official documentation: https://huggingface.co/docs/lerobot/installation If you’re training / fine-tuning, you typically call forward(...) to get a loss and then: - -policy.chunksize=... - -policy.nactionsteps=... - -policy.maxactiontokens=... - -policy.gradientcheckpointing=true You can use the record…
Paper: Zheng et al., 2025, “X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model” (arXiv:2510.10274) Successful generalist Vision-Language-Action (VLA) models rely on effective training across diverse robotic platforms…
Parameters880M
Context—
Weights3.5 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads10.6k
Runs On
What it takes to serve X-VLA-Pt (880M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 1.8 GB | 2.1 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 0.9 GB | 1.1 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.4 GB | 0.5 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.
Model Card
By Jinliang Zheng, published under apache-2.0, revision c1c4a64a7e03.
Paper: Zheng et al., 2025, “X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model” (arXiv:2510.10274) Successful generalist Vision-Language-Action (VLA) models rely on effective training across diverse robotic platforms with large-scale, cross-embodiment, heterogeneous datasets. To facilitate and leverage the heterogeneity in rich robotic data sources, X-VLA introduces a Soft Prompt approach with minimally added parameters: we infuse prompt-learning concepts into cross-embodiment robot learning, introducing separate sets of learnable embeddings for each distinct embodiment. These embodiment-specific prompts empower VLA models to exploit cross-embodiment…
Read Jinliang Zheng's full model card
X-VLA 0.9B (Foundation Edition)
Repository: 2toINF/X-VLA
Authors: 2toINF | License: Apache 2.0
Paper: Zheng et al., 2025, “X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model” (arXiv:2510.10274)
Overview
Successful generalist Vision-Language-Action (VLA) models rely on effective training across diverse robotic platforms with large-scale, cross-embodiment, heterogeneous datasets. To facilitate and leverage the heterogeneity in rich robotic data sources, X-VLA introduces a Soft Prompt approach with minimally added parameters: we infuse prompt-learning concepts into cross-embodiment robot learning, introducing separate sets of learnable embeddings for each distinct embodiment.
These embodiment-specific prompts empower VLA models to exploit cross-embodiment features effectively. Our architecture—a clean, flow-matching-based VLA design relying exclusively on soft-prompted standard Transformers—achieves superior scalability and simplicity.
Trained on Bridge Data and evaluated across six simulations and three real-world robots, the 0.9B-parameter X-VLA simultaneously achieves state-of-the-art performance across diverse benchmarks, demonstrating flexible dexterity and fast adaptation across embodiments, environments, and tasks.
Project Website: https://thu-air-dream.github.io/X-VLA/
Usage
Load the model
from transformers import AutoModel
model = AutoModel.from_pretrained(
"2toINF/X-VLA-WidowX",
trust_remote_code=True
)
Start FastAPI server
from transformers import AutoProcessor
processor = AutoProcessor.from_pretrained("2toINF/X-VLA-WidowX", trust_remote_code=True)
model.run(processor, host="0.0.0.0", port=8000)
Client-server evaluation
You can run the provided evaluation client from our GitHub: 2toINF/X-VLA – Client & Server Code
Architecture
| Component | Role |
|---|---|
| Florence 2 Encoder | Vision-Language representation backbone (encoder-only). |
| SoftPromptedTransformer | Flow-matching action denoiser using learnable soft prompts per embodiment. |
| Action Hub | Defines action spaces, masking rules, pre/post-processing, and losses. |
Training Summary
| Setting | Value |
|---|---|
| Training Data | Heterogeneous Datasets |
| Parameters | ≈ 0.9 B |
| Action Mode | ee6d |
| Precision | BP16 |
| Framework | PyTorch + Transformers |
License
Copyright 2025 2toINF (https://github.com/2toINF)
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
http://www.apache.org/licenses/LICENSE-2.0
Citation
@article{zheng2025x,
title = {X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model},
author = {Zheng, Jinliang and Li, Jianxiong and Wang, Zhihao and Liu, Dongxiu and Kang, Xirui
and Feng, Yuchun and Zheng, Yinan and Zou, Jiayin and Chen, Yilun and Zeng, Jia and others},
journal = {arXiv preprint arXiv:2510.10274},
year = {2025}
}
Links
- Paper: arXiv 2510.10274
- Code & Client/Server: GitHub – 2toINF/X-VLA
- Model Hub: Hugging Face – 2toINF/X-VLA-0.9B-WidowX
Configuration
- Architecture
- XVLA
- Hidden size
- 1,024
- Model type
- xvla
Identity and Version
- Repository
- 2toINF/X-VLA-Pt
- Publisher
- Jinliang Zheng
- Task
- Robotics
- Modality
- Control
- Library
- Not stated by the source
- Parameters
- 880M parameters
- Languages
- vla
- Revision
- c1c4a64a7e03ac5b95c468bf1578f3d03651b53b
- First published
- 2025-11-04
- Last updated
- 2025-11-16
Files and Weights
15 files, 3.5 GB in total. The weights are 1 file totalling 3.5 GB in safetensors.
Weights1 file · 3.5 GB
Configuration9 files · 192.2 KB
Tokenizer3 files · 2.5 MB
Documentation1 file · 4.5 KB
Repository1 file · 1.5 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 3.5 GB | 433acffc992f |
| action_hub.py | Configuration | 9.9 KB | — |
| config.json | Configuration | 2.7 KB | — |
| configuration_florence2.py | Configuration | 15.1 KB | — |
| configuration_xvla.py | Configuration | 3.4 KB | — |
| modeling_florence2.py | Configuration | 127.5 KB | — |
| modeling_xvla.py | Configuration | 11.7 KB | — |
| preprocessor_config.json | Configuration | 766 B | — |
| processing_xvla.py | Configuration | 7.5 KB | — |
| transformer.py | Configuration | 13.7 KB | — |
| README.md | Documentation | 4.5 KB | — |
| .gitattributes | Repository | 1.5 KB | — |
| tokenizer.json | Tokenizer | 1.4 MB | — |
| tokenizer_config.json | Tokenizer | 34 B | — |
| vocab.json | Tokenizer | 1.1 MB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 3.5 GB
Download from Jinliang Zheng
Released by Jinliang Zheng through its official repository on Hugging Face. Read the license.
Built From
- Derived from microsoft/Florence-2-large
- Described by arXiv:2510.10274
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 3.5 GB |
| 16-bit | 1.8 GB |
| 8-bit | 0.9 GB |
| 4-bit | 0.4 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About X-VLA-Pt
How much GPU memory does X-VLA-Pt need?
About 2.1 GB at 16-bit and 0.5 GB at 4-bit: the weights (880M parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run X-VLA-Pt on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use X-VLA-Pt commercially?
Yes. X-VLA-Pt is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
Similar Models
Paper: Zheng et al., 2025, “X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model” (arXiv:2510.10274) Successful generalist Vision-Language-Action (VLA) models rely on effective training across diverse robotic platforms with large-scale, cross-embodiment, heterogeneous datasets. To facilitate and leverage the heterogeneity in rich robotic data sources, X-VLA introduces a Soft Prompt approach with minimally added parameters: we infuse prompt-learning concepts into cross-embodiment robot learning, introducing separate sets of learnable embeddings for each distinct embodiment. These embodiment-specific prompts empower VLA models to exploit cross-embodiment…
SmolVLA is a compact, efficient vision-language-action model that achieves competitive performance at reduced computational costs and can be deployed on consumer-grade hardware. This policy has been trained and pushed to the Hub using LeRobot. See the full documentation at LeRobot Docs. For a complete walkthrough, see the training guide. Below is the short version on how to train and run inference/eval: Writes checkpoints to outputs/train/ /checkpoints/. Prefix the dataset repo with eval\ and supply --policy.path pointing to a local or hub checkpoint.
Flash Vision-Language-Action Inference for Autonomous Driving DFlash draft model for z-lab/Alpamayo-1.5-10B, used by FlashDrive to accelerate the chain-of-causation reasoning of Alpamayo 1.5. DFlash (ICML 2026) uses a lightweight block-diffusion draft to propose several tokens in parallel; the target verifies each block in a single forward, preserving its output distribution. This draft is a 2-layer Qwen3-style network (block size 8) conditioned on target hidden states from layers 24/30/31/32/34. The repository also ships maskembedding.pt, the trained mask-token embedding FlashDrive appends to the target's embedding table. See the base model card and the FlashDrive repository for the full…
SmolVLA is a compact, efficient Vision-Language-Action (VLA) model designed for affordable robotics, trainable on a single GPU and deployable on consumer hardware, while matching the performance of much larger VLAs through community-driven data. Original paper: (SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics)[https://arxiv.org/abs/2506.01844] For full installation details (including optional video dependencies such as ffmpeg for torchcodec), see the official documentation: https://huggingface.co/docs/lerobot/installation If you’re training / fine-tuning, you typically call forward(...) to get a loss and then: - -policy.chunksize=... - -policy.nactionsteps=...…
SmolVLA is a compact, efficient vision-language-action model that achieves competitive performance at reduced computational costs and can be deployed on consumer-grade hardware. This policy has been trained and pushed to the Hub using LeRobot. See the full documentation at LeRobot Docs. For a complete walkthrough, see the training guide. Below is the short version on how to train and run inference/eval: Writes checkpoints to outputs/train/ /checkpoints/. Prefix the dataset repo with eval\ and supply --policy.path pointing to a local or hub checkpoint.