WALL-OSS (flow)
X Square Robot, China · September 2025
- Parameters
- 4 B
- The Hugging Face card lists 4B parameters.
- 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 is X Square Robot's first open-source embodied foundation model, a 4B parameter system built on a Qwen2.5-VL-3B backbone. Its architecture uses a statically routed mixture of experts, with a vision-language FFN and an action FFN tightly coupled at every layer, trained first with FAST-tokenized discrete actions and then with continuous flow matching. The paper's central claim is that this avoids catastrophic forgetting, and the embodied VQA table supports it: object grounding at 91.6 percent against 46.1 percent for the untouched Qwen2.5-VL-3B backbone. The Block-Spell results are the more useful signal for practitioners, showing that co-training on VQA and actions lifts fine-grained instruction following from 26 to 87 percent while pi-0 action-only reaches 9 percent. It is Apache-2.0 and LeRobot-integrated, but publishes no chunk size, control rate, VRAM or latency figures.
Architecture
- Backbone
- Qwen2.5-VL-3B
- Action head
- Mixture-of-experts with static routing (not learned softmax or top-k): a tightly coupled Vision-Language FFN and a separate Action FFN at each layer. Trained in two stages, an Inspiration stage with discrete actions via FAST tokenization followed by an Integration stage with continuous actions via flow matching.
- Parameters
- The Hugging Face card lists 4B parameters. The paper names QwenVL2.5-3B as the backbone, so roughly 1B sits in the action pathway. Model class is Qwen2_5_VLMoEForAction, BF16.
- Pretraining data
- The paper states the overall corpus exceeds tens of thousands of hours. It combines self-collected robot action data from desktop arms, mobile stands, wheeled bi-arm systems and wheeled humanoids, open-source action data from more than 24 datasets including DROID, BC-Z and BridgeData, and multimodal VQA for general perception and embodied spatio-temporal reasoning.
- Embodiments
- desktop arms, mobile stands, wheeled bi-arm systems, wheeled humanoids, x2_normal configuration referenced in the released configs
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.
| Precision | Weights | Note |
|---|---|---|
| fp32 | 15 GB | Training master weights, rarely used for inference |
| bf16 or fp16 | 7.5 GB | The usual way these checkpoints are served |
| int8 | 3.7 GB | Quantised, expect some loss of precision on fine motions |
| int4 | 1.9 GB | Aggressive 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.
Neither the paper nor the repository publishes a VRAM figure for this variant. The stack requires CUDA 12.x, Python 3.10, FlashAttention 2.8.3 and Ubuntu 22.04, and training uses FSDP, which implies a multi-GPU setup for full fine-tuning but is not a stated number.
The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.
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
No results in this category are published for this model.
Real world
No results in this category are published for this model.
Fine tuned tasks
No results in this category are published for this model.
Embodied reasoning
Offline benchmarks that test understanding rather than control. No robot moves in these, so a high score says nothing about grasping.
- Real robot Block-Spell (WALL-OSS paper, Table 3) number blocks, fine-grained instruction following, VQA co-trained95%success rateAuthors' own evaluation. Action-only scores 80 percent and pi-0 action-only 35 percent.source
- Embodied VQA (WALL-OSS paper, Table 2) object grounding91.6%accuracyAuthors' own evaluation. The untouched Qwen2.5-VL-3B backbone scores 46.1 percent, which is the evidence for the no-forgetting claim.source
- Embodied VQA (WALL-OSS paper, Table 2) scene captioning87.6%
- Real robot Block-Spell (WALL-OSS paper, Table 3) letter blocks, fine-grained instruction following, VQA co-trained87%success rateAuthors' own evaluation. The same model trained action-only scores 26 percent and pi-0 action-only scores 9 percent, so this row measures the training recipe as much as the architecture.source
- Embodied VQA (WALL-OSS paper, Table 2) action planning69%accuracyAuthors' own evaluation. Qwen2.5-VL-3B baseline is 59.8 percent, so the margin here is much smaller than on grounding or captioning.source
On real hardware
Real-robot evaluations on Set-Table, Tidy-Bedroom and Block-Spell tasks, reported as outperforming pi-0 and Diffusion Policy. The most informative real result is the Block-Spell instruction-following table, where co-training on VQA plus actions lifts letter-block accuracy from 26 percent (action-only) to 87 percent, against 9 percent for pi-0 action-only.
Fine tuning it yourself
Data is prepared in LeRobot format and the repository integrates with LeRobot directly. Training uses FSDP via a train_fsdp entry point, with a configurable robot DOF setting per embodiment. Inference runs in validate mode at bfloat16. The Apache-2.0 licence permits commercial use, which distinguishes it sharply from GO-1.
Where it helps, where it does not
Strengths
- Apache-2.0 on both code and weights, so commercially usable, which is the main practical advantage over GO-1
- Keeps genuine vision-language competence after action training: 91.6 percent object grounding against 46.1 percent for the raw Qwen2.5-VL-3B backbone, which is direct evidence against catastrophic forgetting
- Compact 3B backbone, 4B total, which is at the small end of this comparison
- Ships both a flow-matching and a FAST-token variant from the same architecture, so the discrete-vs-continuous tradeoff can be tested without changing stacks
- Integrated into the LeRobot framework, which lowers the data-plumbing cost
Limits
- No published action chunk size, control frequency, VRAM figure or latency for this variant, so deployment sizing has to be measured, not looked up
- Real-robot manipulation results appear as bar figures rather than a numeric table, so the claimed margin over pi-0 on Set-Table, Tidy-Bedroom and Block-Spell cannot be quoted precisely
- Action-only training collapses instruction following: 26 percent on letter blocks versus 87 percent with VQA co-training, so the recipe, not just the checkpoint, is load-bearing
- The Hugging Face card carries no YAML metadata and no licence field; the Apache-2.0 licence is established from the GitHub repository
- Superseded by Wall-OSS-0.5 in May 2026
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://arxiv.org/abs/2509.11766
- https://arxiv.org/html/2509.11766v1
- https://huggingface.co/x-square-robot/wall-oss-flow
- https://github.com/X-Square-Robot/wall-x
- https://github.com/X-Square-Robot/wall-x/blob/main/README.md
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.