XSR

Wall-OSS-0.5

X Square Robot, China · May 2026

Foundation VLAOpen weights, Apache-2.0
Parameters
4 B
4B parameters total on a 3B vision-language backbone (Qwen2.5-VL-3B-Instruct), so roughly 1B in the action expert.
GPU memory
7.5 GB
weights at bf16, computed. Fits a 12 GB card
Inference latency
not published
per action step
Weights
Apache-2.0
downloadable checkpoint

What it is

Wall-OSS-0.5 is X Square Robot's May 2026 successor to WALL-OSS, a 4B parameter model on a 3B Qwen2.5-VL backbone using a Mixture-of-Transformers layout with separate vision-language and action pathways at every layer. The technical report is built around a question most VLA papers avoid, namely whether pretraining alone produces usable robot behaviour or only a better fine-tuning initialization, and answers it with a 17-task zero-shot suite reaching 51.1 percent average task progress before any adaptation. After fine-tuning it reaches 60.5 percent average task progress on 15 real tasks against 43.0 percent for pi-0.5. It is one of the very few models here that publishes a control rate on named hardware, at 15 Hz on an RTX 5090 at 448x448 with 10 denoising steps. The metric to watch is that all headline real-robot numbers are task progress rather than binary success.

Architecture

Backbone
Qwen2.5-VL-3B-Instruct
Action head
Mixture-of-Transformers (MoT) with a separate VL Expert and Action Expert pathway at each layer, so modality-specific routing with end-to-end gradient flow. Trained with gradient-bridged co-training that combines discrete action prediction, multimodal prediction and continuous flow matching. Inference uses 10 denoising steps by default. Linear projections including QKV and output projections are not shared between action tokens and text or vision tokens.
Control rate
15 Hz
Parameters
4B parameters total on a 3B vision-language backbone (Qwen2.5-VL-3B-Instruct), so roughly 1B in the action expert. Stated in both the technical report and the Hugging Face card.
Pretraining data
More than 1 million robot trajectories per epoch across more than 20 distinct embodiments, roughly 60 percent self-collected and 40 percent open-source, plus about 90 million multimodal samples (78M open-source and 12M embodied bridge samples). Open-source components include DobbE, RH20T, UMI-biarm, BridgeData V2, DROID, FurnitureBench and RoboMind, RoboChallenge and RoboCoin series.
Embodiments
proprietary ex_normal and x2_normal configurations, more than 20 embodiments in pretraining, covering 24 open-source dataset configurations

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 4 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
fp3215 GBTraining master weights, rarely used for inference
bf16 or fp167.5 GBThe usual way these checkpoints are served
int83.7 GBQuantised, expect some loss of precision on fine motions
int41.9 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 12 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.
Control rate15 Hz

No VRAM figure is published. The only concrete hardware datapoint is that inference was measured on an RTX 5090, which places the model within a 32GB consumer card, but the report gives no memory number so none is recorded.

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.

  • LIBERO average, action-space-supervision ablation
    96.5%
    success rateAuthors' own evaluation, reported as a peak average inside an ablation comparing action-space against velocity-space supervision, not as a primary benchmark claim.source

Real world

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

  • Real robot, 15-task suite (Wall-OSS-0.5 report) manipulation subset, 10 tasks, after fine-tuning
    61.1%
    average task progressAuthors' own evaluation.source
  • Real robot, 15-task suite (Wall-OSS-0.5 report) after fine-tuning, overall average
    60.5%
    average task progressAuthors' own evaluation. pi-0.5 scores 43.0 percent on the same suite, a 17.5 percentage point gap. Task progress, not binary success rate.source
  • Real robot, 15-task suite (Wall-OSS-0.5 report) reasoning subset, 5 tasks, after fine-tuning
    59.3%
    average task progressAuthors' own evaluation. Close to the manipulation subset, which the authors read as evidence that reasoning ability survives action training.source

Fine tuned tasks

What the model reaches after being adapted to a new task or a new robot, usually from a small number of demonstrations. This is the number that matters if you bring your own data.

  • Real robot, 17-task zero-shot suite (Wall-OSS-0.5 report) zero-shot at the 400k pretraining checkpoint, overall average
    51.1%
    average task progressAuthors' own evaluation. Split into 50.0 percent on seen tasks and 53.6 percent on unseen tasks. This is the paper's core evidence that pretraining alone yields non-trivial behaviour.source

On real hardware

Evaluated on a 17-task zero-shot suite and a 15-task fine-tuned suite on real hardware. Zero-shot at the 400k checkpoint reaches 51.1 percent average task progress (50.0 percent seen, 53.6 percent unseen), with Block Sorting at 100 percent, Fruit Sorting at 96 percent, Ring Stacking at 86 percent and Rope Tightening at 82 percent. After fine-tuning, 60.5 percent average task progress on 15 tasks against 43.0 percent for pi-0.5. Runs at 15 Hz on an RTX 5090.

Fine tuning it yourself

Same wall-x stack as earlier WALL-OSS releases, with Wall-X 1.1.0 (June 2026) adding a public serving and evaluation runtime plus DMuon training support for this checkpoint. Data goes through LeRobot format. Action space is 26-dimensional (20D bimanual end-effector plus gripper, 3D mobile base, 1D lift, 2D head) with 6D rotations and relative poses, and the action horizon is fixed at 1 second with the frame count adapted to each platform's control rate.

Where it helps, where it does not

Strengths

  • Publishes a real control frequency on named consumer hardware (15 Hz at 448x448, 21 Hz at 224x224 on an RTX 5090), which almost no competing VLA does
  • Apache-2.0, commercially usable, at 4B parameters
  • Reports zero-shot task progress before any fine-tuning (51.1 percent over 17 tasks), which directly answers whether pretraining alone yields usable behaviour rather than only a better initialization
  • Beats pi-0.5 by 17.5 percentage points after fine-tuning on the same 15-task suite (60.5 versus 43.0)
  • Wide action space covering bimanual arms, mobile base, lift and head, so it addresses whole-robot control rather than arms alone
  • Pretrained across more than 20 embodiments with 1M-plus trajectories per epoch

Limits

  • Task progress, not binary success rate, is the headline metric on both real-robot suites, so 60.5 percent does not mean 60.5 percent of tasks completed
  • Absolute zero-shot performance is around 51 percent task progress, which is far from deployable without fine-tuning despite the pretraining scale
  • No VRAM figure is published, only the GPU model used for the speed measurement
  • The 15 Hz figure depends on 448x448 input and 10 denoising steps; changing either changes the rate, and 15 Hz is low for contact-rich or dynamic tasks
  • Real-robot evaluation is on X Square Robot's own platforms and its own task suites, with no third-party reproduction
  • LIBERO results appear only inside an action-space-supervision ablation rather than as a headline comparison

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.