WALL-OSS-0.5

WALL-OSS-0.5 is an open-source vision-language-action (VLA) model developed by X Square Robot (自变量机器人, Shenzhen) and released on 28 May 2026. It belongs to the WALL family of foundation models and is an evolution of the earlier WALL-OSS. Its motto is “Pretrain Once, Act Anywhere” — the model is deployment-ready straight from the pretrained checkpoint, with no per-task fine-tuning required.

The model has about 4 billion parameters and is built on a 3-billion-parameter vision-language model (VLM) backbone augmented with action-generation components. During training, discrete action tokens shape the VLM backbone; at inference, continuous flow matching supplies the action output; and grounded multimodal data preserves the vision-language prior (gradient-bridged co-training).

WALL-OSS-0.5 was trained on data from more than 20 robot form types, millions of trajectories, and a multimodal corpus of about 90 million items. It takes vision, sensor and robot-state data plus natural-language instructions as input, and produces continuous control actions as output. The model achieves non-trivial zero-shot results on a 17-task suite and a 60.5% average task progress on 15 real-robot tasks (after fine-tuning).

The model is released under the Apache-2.0 license together with weights, code and tooling (GitHub, Hugging Face). This permits commercial and non-commercial use, self-hosting, and fine-tuning across different robot embodiments.